In May of 2000 I signed a one-page contract with a company called ReviewNet and became, in their words, a “Technology Review Team Member.” Five hundred dollars a month, work-for-hire, everything I wrote theirs forever. What I wrote was tests. Not school tests — hiring tests. ReviewNet sold employers a bank of technical screening questions, and a candidate who wanted a programming job would sit down and answer mine before a single human at the company ever gave them a chance. I’d started a couple of years earlier with a fifty-question Java exam. By 2000 I was writing and vetting questions across the whole stack a web shop hired for, Java fundamentals up through J2EE and JSP and the scripting glue that held it together. I was, in the most literal sense the language allows, a gatekeeper. I built the turnstile and never had to stand at it.

Why the gate existed

Before I either defend or indict the thing, you have to see the hole it was filling, because the hole was real and it still is. Software development has no working credential. A computer science degree certifies that you can reason about algorithms, which is not the same as being able to ship a functioning piece of software with real features, and everyone who has ever made a hire knows it. Some of the best developers I’ve worked with never finished college. Some of the worst held advanced degrees. There’s no bar exam, no medical board, no license to practice. Microsoft had been certifying people since 1993, and the certs tested narrow product trivia that often had little to do with whether you could actually build anything. The industry knew it, too: by the late nineties it had a name for the graduates — paper MCSEs , certified on the exam and lost in a server room — and a joke about what the letters really stood for, Minesweeper Consultant and Solitaire Expert. The one credential the field had, the field didn’t believe. So a hiring manager looking at four hundred résumés had, functionally, nothing — no accreditation to lean on, no signal to trust, and a real job to fill.

And the obvious alternative — sitting your best engineers down to interview every plausible candidate — cost the one thing a software shop can least spare: those same engineers’ hours at the keyboard. An automated question bank scaled. It promised to vet people without pulling the people you’d hire them to work alongside off their own work. That promise is most of why the thing existed at all.

ReviewNet sold that manager a signal. A crude one, but a signal. That’s the honest origin of the thing. Not a scam dreamed up to charge tolls, but a primitive first swing at a genuinely hard, genuinely unsolved problem: how do you vet competence in a field that refuses to be credentialed? Call it version 0.1. I wrote a great deal of version 0.1.

From the wrong side of it

I know that vacuum from the receiving end, which is the part that makes the rest of this hard to write. Long before ReviewNet, I came out of an electrical engineering program with something close to a 2.9 GPA. Engineering school was absolute survival for me. I kinda knew at year three that I really should’ve chosen computer science but I was determined to see it through, and I did…with that 2.9. My wife went through the same kind of program with a 4.0 and walked straight into a job at Motorola; when she hand-carried my resume in, they wouldn’t take the application. Not “no thanks” — they wouldn’t let me in the door far enough to be turned down on the merits. The filter ran on a number, my number was below the line, and no human was ever going to see the rest of me.

What the number couldn’t see was that I’d spent college actually building things — shipping working software to real, paying customers, the messy applied work that barely moves a GPA. By the one measure that turns out to matter on the job, I was further along than my transcript said. The software world was full of people like that: self-taught, half-credentialed, dropouts and B-students who could nonetheless build the thing, diamonds in the rough if you knew how to spot us, invisible if you didn’t. A system built around accreditation had no way to see us at all. Mostly still doesn’t.

Then I grew up and built the filter to filter out future me’s. Ouch.

What it actually measured

The trouble with the version 0.1 hiring filter was the trouble with any test made of questions. It couldn’t measure whether you could do the job; it could only measure whether you could answer questions about what the job requires, and those are different skills. Anyone who has crammed for a test and forgotten it by Monday knows they’re different.

Fifty Percent of What? made the case that we keep reaching for a number because the things we actually fear are hard to measure and a number at least looks countable. A proxy. The screening test is that same reflex aimed at hiring. Can this person do the work is expensive to find out; you have to hand them real work and watch. So we sold a cheaper thing in its place — fifty questions standing in for an interview standing in for the job. The bet was that someone who could answer them off the top of his head had probably done the reps, and someone who’d done the reps could probably do the work. Probably. The whole instrument ran on that “probably.”

And here’s the deeper flaw, the one only hindsight makes obvious: rote recall of syntax and API trivia turns out to have almost nothing to do with whether a person can actually build software. The best developers I know don’t carry encyclopedic API details in their heads at all; they have the opposite skill — reading an unfamiliar interface fast, recognizing the pattern under it from the hundred they’ve seen before, sensing where the gotchas are buried. It’s what Tom and Ray Magliozzi did on Car Talk for thirty-five years: diagnose a car they couldn’t see, on the other end of a phone line, by asking the right handful of questions. Neither of them was reciting part numbers. They had the manuals for that.

The man who can fix it is on the other end of that line. He has never seen this engine and never will. He'll find it anyway.

Same in software. The competence was never the memorized fact. It was knowing how to find and absorb the fact you didn’t have, which in 2000 meant knowing your way around the documentation in the context of creative problem solving. Which is the joke at my expense, and the whole essay in one line: ReviewNet hired me to write these tests partly because I’d written the reference books on the subjects — and then we built a test that pretended the reference didn’t exist.

Where the machine actually changes things

Here I have to be careful, because the easy takeaway from this is wrong. The easy version says a language model answers all fifty hiring questions in seconds, so the test is dead. But plenty of companies didn’t use these as take-home screens; they sat the candidate at a machine in the office and watched — proctored, live, the answers coming out of his own head or not at all. In that room the model on your phone does you no good.

What AI changed is quieter and worse. The thing my questions rewarded was unaided recall, and that is now exactly the part of the job a machine does for free. On any real workday the developer I was screening will have the model right at his elbow, the way he’ll have the docs and the search bar and his own old code. The way the Magliozzis had Chilton’s automotive reference manuals . So even a proctored test that runs perfectly is measuring the candidate in the one condition he will never actually work in: alone, from memory. You can run it flawlessly and still learn the wrong thing, because you tuned it for a skill the job stopped requiring the day the machine arrived at every desk.

The one I was proud of

There was one piece of it I still kinda like, because it was the only part trying to be a forge instead of a tollbooth , and because the instinct behind it was right. Call it version 0.2. In 2002 I built what we called a scenario: a small Java program presented as a running thing instead of a multiple-choice list, a slideshow walking the candidate through what it did, with questions that asked not what does this keyword mean but what would you do here, and why. The idea was to quit testing recall and start testing judgment: watch someone reason through a real-ish situation instead of reciting facts.

That instinct didn’t die. It won. It’s essentially what the famously challenging interviews at places like Google do now — put a person in front of a problem and watch them think, live, out loud. My version was a dial-up cartoon of a thing the whole industry eventually agreed was the right idea, so I don’t get to be smug about how crude the fifty questions were; the better answer was already visible in 2002, and serious companies spent the next two decades building it out.

And version 0.2 is beatable too, on the same terms — if you let the candidate use a model. Which exposes the real question hiding under the whole enterprise: are you measuring what a person can do alone, or what they can do with the tools they’ll actually have on the job? That’s a question about taste rather than memory, and it’s the one every test now has to answer for itself. This one was about the gate, and the gate I built measured none of it.

The part that’s on me

What I’m least proud of isn’t the crudeness of the thing. It’s the overclaim. Everyone in that business knew the test was a “probably,” and the business needed it sold as a “definitely,” and the gap between those two words is where both the money and the unfairness lived. Good people got filtered out by my questions for the crime of not having memorized something they could have looked up in twenty seconds. I knew exactly what that felt like. I’d been filtered out myself, by a different number, and I built the machine anyway. For five hundred dollars a month.

What’s left

So when people ask whether AI is going to wreck technical hiring, I think they’re braced for the wrong impact. The screening gate was never a sturdy thing the machine knocked over. It was a version-0.1 patch over a hole the industry has never fixed — no reliable way to certify a developer — and AI didn’t make the hole. It just flooded the desk with the exact capability the tests were built to ration, until the patch looked as flimsy as it always was. And now we’re seeing tech hiring devolve into machines reviewing documents created by machines…hardly an improvement.

But pull off the recall test and the scenario test and every version that comes after, and you reach the thing none of them could touch — the thing my four years were a way of not admitting. You can’t really know a person from the outside. Technical skill was always just the foot in the door, the cheapest and most measurable part, which is exactly why we measured it. Everything that actually decides whether a hire works is intangible: judgment, what someone does when it breaks at two in the morning, whether they finish what they start, whether they write horrific spaghetti code, whether you’d want them in the room on the worst day. No question bank reaches that. No scenario reaches it. The best interview at the best company is still, under the suit it’s wearing, a guess.

The only thing that ever actually worked for me was the slow way. A family member once suggested I take up golf, for the networking, for business. It struck me as the worst reason anyone has ever picked up an activity. I skateboarded instead, because I loved it, and some of the longest working relationships of my life came out of that. The job I have right now traces back to a friendship built on a board when I was a teenager. Stephen, who I still work with, once got asked how he finds good people, and he answered — joking, but not really — “well, it’s sometimes people I skateboarded with at my grandmother’s house.” Nobody was networking. We were just there, for years, doing a hard thing with each other and taking our lumps along the way, which turns out to be the most you can ever learn about a person.

That’s the joke at the bottom of the whole business, mine included. Hiring is a crapshoot we dress up as a process, because the honest alternative — that you can’t know a stranger until you’ve worked beside them — doesn’t scale and doesn’t sell. Years of someone’s public record can show you a great deal; a résumé caught at the gate, on a Tuesday, shows you almost nothing. The test lives at the gate, and the gate is the worst possible place to try to know anyone. A number that couldn’t see what I’d built kept me out of a job I could already do; I spent four years building numbers with the same blind spot. And the machine that finally exposed the tests can’t patch what’s underneath, because what’s underneath was never a measurement problem. It’s the oldest one there is: another person, and how little of them you can know before you decide to trust them. The honest move was always to hand someone something real and find out together.

In The Diamond Was Always a Story , I argued that the diamond’s premium was a story, and that the lab-grown stone didn’t destroy it so much as expose it. The value was never in the rock; it was in a sentence someone wrote and a scarcity someone arranged. Then I widened it. Every livelihood is a rent on a scarcity — the scribe’s patient literacy, the single wage that once bought a house, the fifteen-year window when there was more attention around than decent stuff to fill it. The rent lives in the gap between what people want and who can supply it, and when the gap closes the rent dies, however good the work was on the way out.

So this one sorts the rents. They are not all the same kind of thing, and which kind you’re holding is everything once the flood arrives — the flood of cheap, competent, machine-made work now rising against all of them. I want to turn that question away from the jewelry case, toward the gallery, the imprint, the festival, the playlist — the whole apparatus that decides which creative work gets to be worth something.

Forges and tollbooths

Some creative premiums are forges, where the price tracks something a person actually built: earned judgment, a sensibility you can’t fake, the formation that only comes from doing the thing badly for years until you could do it well. Forges look like they’re about craft. They’re about the judgment those years of craft were quietly building. That’s a scarcity too — everything is — but it’s the deepest kind, the last thing a flood reaches, because the thing being paid for was never on the surface of the artifact. A perfect copy of the output doesn’t reproduce it.

The cleanest example is the one that looks least like craft. Rick Rubin has produced some of the most important records of the last forty years, and he can’t play an instrument, doesn’t run the board, reads no music — his entire contribution is taste. On the Beastie Boys’ 1986 Licensed to Ill it was welding rock riffs onto hip-hop beats. Less than a year later with The Cult it was hearing past a finished, $360,000 album the band had already soured on — “soup,” the singer called it — and re-cutting the same songs raw until they became Electric. Same band, same songs; the one variable was his judgment. That’s the forge in its purest form: the ear that knows what a record wants to be. Real, earned, and genuinely missed when it goes.

1986. Rubin's gag on rockstar excess — hand the kids their own private jet, then fly it into a mountain. He couldn't have drawn it; he just knew it should crash.

Other creative premiums are tollbooths. The price tracks access untethered from ability. Somebody stands at a narrow place on a busy road, controls who gets through, and the markup is rent on the passage, not payment for the work. The work might be good or it might be mediocre; the tollbooth doesn’t especially care, because what it’s selling is the scarcity of having been let through.

Rubin sharpened that ear inside one of these. The label he co-founded with Russell Simmons, Def Jam, was a tollbooth the way every record label was: it didn’t make the music better, it controlled who got pressed, promoted, and played. That cut was rent on access to being heard — the forge and the toll in a single company.

AI is the lab diamond for the tollbooth kind of premium. It floods the narrow place with competent, identical-looking output until waiting in line at the gate stops meaning anything, and it forces the question every tollbooth spent decades making sure you’d never ask. Was I paying for the work, or was I paying for the access? For the forges, the flood can match the execution — even a master’s — and still miss what you were paying for, which was never in the artifact: the judgment that made it, and will make the next call too. Part of that judgment is timing — being first to hear the thing everyone will later find obvious. Once the first ear proves the sound, the imitators rush the door; they have the move, not the ear that found it, and it has already moved on. For the tollbooths, it’s 1947 all over again — the year a copywriter promised us a diamond is forever — and the slogan is now wearing off.

A field guide to the gates

So let’s walk the forge/tollbooth apparatus and sort it, because the whole argument lives in the sorting.

The anointing ladder — art school to agent to gallery to prestige imprint — sells the right to be counted as a real artist. Its defenders will tell you it protects quality, and sometimes it did: a great editor is a forge, the Rubin kind of judgment aimed at a manuscript instead of a mixing board, and it’s real. But it was always the small part of what the ladder charges for. Strip it out and what’s left is mostly a tollbooth: a sorting mechanism that converted the scarcity of its own approval into a premium on everyone it approved. AI doesn’t flood the editing. It floods the approval, by making competent work cheap enough that the gate can no longer pretend its stamp was the same thing as the quality.

Prestige scarcity (the gallery wall, the numbered edition, the velvet rope) is closer to pure De Beers — the diamond cartel that made a common stone feel rare by sitting on the supply. It staples artificial scarcity onto work that is often infinitely reproducible and charges for the staple. Supreme turned that into an empire — a forty-dollar cotton box-logo tee that resells for hundreds, not because the cotton or even the logo is particularly special but because they made only so many and dropped them on a Thursday. Same move as the diamond, a century on, in streetwear. This one was always a tollbooth wearing a forge’s clothes, and it’s the one I’ll shed the fewest tears for.

Gatekept distribution — shelf space, playlist placement, the festival slot, the residency — is a tollbooth the internet already half-dissolved and AI finishes off. And a lot of what reads as merit here was never merit: the table of new hardcovers at the front of the bookstore isn’t the booksellers’ favorites, it’s co-op placement the publisher paid for, and most of us go a whole lifetime thinking that table means good. I had books in those stores once, well back from that table, and I would never have quietly turned one of mine face-out when no one was looking. Not once. When a few people controlled the few channels, getting through them was worth a premium because getting through was the whole game. The channels multiplied; the toll kept charging anyway, on momentum.

The signature markup is the slipperiest gate, and Jeff Koons is standing in the middle of it. Koons is one of the most expensive living artists in the world. His signature works are giant balloon animals in mirror-polished steel, and one of the dogs sold for $58 million. He also hasn’t personally fabricated most of his art in decades — at his peak he ran a 16,000-square-foot studio where as many as a hundred and twenty assistants executed his work on a color-by-numbers system, so a hundred hands would read as one. A balloon dog is a Koons because he says it is, and the market has never once flinched. That’s the whole puzzle in one man: is the name buying you rare vision you’re glad to keep paying for, or a scarcity that was arranged, authenticated, and about to stop being scarce? Same signature, two completely different things, and the flood is what tells them apart. There’s a sharper blade buried in Koons — that a workshop of anonymous hands producing identical work under one name is the very thing we’re now told to fear about machines, and got called genius when the name was his — but that one gets its own essay. Here he’s just Exhibit A that the name and the work were never the same thing.

The part that costs people I like

It would be cheap to run this as nothing but good news. Tollbooths made us stand behind unnecessary velvet ropes, and AI is cutting those ropes to set us free. But it’s not that simple…there’s a cost.

When a tollbooth collapses, that is a correction, not a tragedy. Nobody is robbed of anything real; a rent that was being charged on access stops being collectible, and the work that was actually good is still actually good.

But the gate falling takes more than the gatekeeper. It takes the people who only ate because the gate rationed supply, and some of them are real makers I’m squarely on the side of. The midlist novelist whose living depended less on being read than on being scarce. The working illustrator whose rate held up because there were only so many of her in the phone book. The scarcity that protected the tollbooth was also, accidentally, the thing that fed some genuine talent, and when it goes, the talent doesn’t automatically get fed some better way. Both of these are true at once and we shouldn’t flatten either into the other. The correction is right. The cost is real. Anyone selling you only one of those is selling you something.

What’s left when the gate is gone

What survives the flood is the same thing the diamond left behind. When the arranged scarcity drains out of a thing, the value doesn’t vanish. It relocates, out of the gate where it was never really sitting and into the maker, where it was the whole time. The forge outlasts the flood precisely because a cheap twin of the output can’t carry it. Taste you can’t counterfeit. Judgment that took a life to build. The specific, answerable sensibility of a person who did the work the hard way and would do it again. None of that was ever in the gate’s gift to grant or withhold. The gate just took credit for it, the way the rock took credit for the diamond.

Which means the skill the flood demands of all of us, buyers and makers both, is the same one this whole publication keeps circling: telling the forge from the tollbooth. Refusing to mourn a toll as if it were a craft, and refusing to wave off a craft as if it were a toll. It’s harder than it sounds, because the tollbooths spent a century dressing as forges and the disguise is good.

There’s a hopeful ending buried in the first example. The Beastie Boys made Licensed to Ill borrowing Rubin’s ear whole — barely an instrument in their own hands on it. Within a few albums they’d picked those instruments back up, taken the controls, and were making records only they could have made. The apprentices grew their craft and built their own ear. That’s the tell that a forge was never a tollbooth: a toll is a gate you keep paying to pass; taste is something you build — badly at first, for years — until one day it’s yours for good.

I started by telling you nobody anointed this. That’s not modesty; it’s the proof of the thing. A few years ago an unvouched-for maker publishing straight to the open road was a curiosity at best, because the gates still rationed who got heard. They ration less every month. I’m not predicting the correction from a safe distance — I’m standing in the middle of it, a guy who got to make a life out of doing the work without ever being let through, which a generation ago wasn’t really on the menu. The gate dissolving is the only reason someone like me gets a say at all. So you’ll forgive me for not grieving it too hard. I’m what comes through when there’s nothing left to charge a toll.

He had it printed.

I had not heard of Trithemius until recently but his complaint I knew well, it’s in my feed every morning wearing new clothes. The abominable machine makes too much, too fast. There’s no soul in it. The people who did the work the hard way are being buried under a flood of cheap product, and something we can’t quite name is slipping away. Every word of that is true right now, about AI. Every word of it was also true in 1492, about type — the scribes had every one of those complaints but the em-dash. That should tell you something, and it isn’t the comfort people usually reach for. It’s better than comfort. It’s an explanation.

Here’s the pattern, plainly, because the whole thing hangs on it. Every livelihood is a rent on a scarcity. Somebody can do a thing most people can’t, or can’t do fast enough, and the gap between what’s wanted and who can supply it is where the money lives. Cristiano Ronaldo’s scarcity is about doing something most of us can’t, and doing it better and longer than almost anyone. For his explosive athleticism, prolific goal-scoring, and unmatched work ethic, he charges the Al-Nassr soccer club north of $200 million a year in rent. The German scribe’s scarcity, while earning a bit less, still had real value for its time. The scarcity was literacy and patience. And the printing press erased it. The scribes were as good as they’d ever been; the craft simply stopped being scarce, and the rent died with it. The same reversal came for portrait painters and switchboard operators and the men who sawed ice out of frozen lakes for a living. It’s happening now, this decade, to a particular kind of person who makes things for a paycheck. I’m one of them.

The window nobody sees

We are terrible at seeing this while it happens, worse still at remembering it afterward. Only two generations ago a family could not only survive but thrive on one income. And not just fancy white-collar incomes with clean hands. Plumbers, electricians, and factory workers alike could support a family with one earner. One wage, a house, kids, a pension waiting at the end. Maybe even a Chevy Bel Air. That world was real, and it was recent, and it is thoroughly gone, and here’s the rub: almost no one can tell you why. Ask around. You’ll get “greed,” or “inflation,” maybe “NAFTA,” or just a shrug. The real answer is a tangle of a dozen threads — a postwar accident of scarcity we mistook for the natural order, a workforce that doubled, inflation, a second paycheck quietly bid away into the price of a house near a good school — and even the economists don’t agree on which thread pulled hardest. That disagreement is the point. When the people paid to explain it can’t, the rest of us were never going to.

That’s what it looks like once a window has closed: a new normal no one can quite explain, filling the space a whole way of life used to occupy. And the space where an explanation should be never stays empty for long. People will take almost any story that names a culprit, because a culprit you can point at — a neighbor, a newcomer, a them — always goes down easier than a tangle of forces with no face. A lot of the anger in any era is grief for a closed window, hunting for a face to blame. Which is the quiet, urgent case for naming the thing plainly: an explanation is what you reach for before someone hands you a scapegoat instead. The people who lived through those changes never saw the shape of them. We might, this time. We’ve watched it happen often enough to know the shape before it’s done. And naming it while it happens beats pretending it isn’t.

Now the one that’s ours

For about fifteen years, call it 2010 to 2025, you could make things — videos, posts, songs, newsletters — put them online, and if enough people liked them, make a living. Not usually a big one. Rent. A car payment. Enough to quit the other job. For a select few, a life-changing amount. We named it the creator economy, and the name did what names do: it turned a temporary arrangement into a permanent-sounding feature of the world, a career as solid as carpenter or nurse. Kids grew up planning to enter it.

It ran on a scarcity almost nobody inside it had to think about. There was more attention in the world than there was decent stuff to fill it, and the platforms were young and starving for supply. So they paid: cash, through ad splits and creator funds, and in kind, by handing your work an audience for free. People inside it thought they were being paid for talent. They were being paid for a bottleneck. Making something watchable still took a person, and a person has a body and a day and a limit, and that limit was the whole market. Your one good video could matter because there were only so many good videos a human could make in a week. That’s scarcity.

AI dissolves the limit. When a machine can turn out a hundred cute dance videos in the time you do one dance, the person’s day stops being the bottleneck, and the bottleneck was the business. You cannot outdance the machines. This is the scribe again, down to the bone. A scribe’s real product was time: the year a human hand needed to reproduce one book. The press didn’t out-write anyone. It made that year worthless. The model is doing the same thing to your day, and the day was what you were being paid for. Not the dance.

So that’s the hard part, no cushion on it. Here’s the part that consoles, and I mean the structural kind, the kind worth knowing.

What the camera did

Go back to 1839, the year the camera arrived in public. There was a whole working class of portrait painters — the jobbing ones, the people who made rent painting your beloved’s face for the parlor wall because a painting was the only way to keep it. Inside a generation the camera did their job faster, cheaper, and truer to life, and the floor fell out from under the portrait painting trade.

1839. The machine that would end the portrait trade, already standing in the studio.

But watch what it did to the painting itself. Once the machine owned accuracy, the painter had no reason left to chase it, and quit. The job of looking exactly like the thing was finished, and finishing it freed the brush for everything else: smear, suggestion, distortion, light, mood, paint that admits it’s paint. The blurry radical work we now file under Impressionism walks straight through the door the camera opened. Painting got fired from a four-hundred-year job and became more itself than it had ever been allowed to be.

That’s the shape worth carrying. When a machine takes the part of a craft that was only ever labor — the exact copy, the reproduced book, the daily quota of watchable stuff — it can free the human part to become what the labor had been crowding out. It doesn’t always. But it can, and it has, and knowing that is worth something while you’re trying to come to terms with machines perhaps displacing the thing you have to offer.

One caveat, because I’ve made that sound smoother than it was. The portrait painter who lost his living in 1845 did not get to enjoy Monet in 1875. He just lost his living, and telling him his ruined trade would one day set French painting free would have been a grotesque thing to say to a man who couldn’t feed his children. Precedent comforts the species, not the person it lands on. My friend Travis Foster has been a commercial illustrator for decades. He already survived one flood — the stock-art boom that gutted commercial illustration in the 2000s, when a nine-dollar download started beating a day’s work — and he did it by reinventing what he sold. Now he’s facing the second, the one that draws from a sentence in seconds, and whether he reinvents again or this is the flood that finally takes him is anyone’s guess. We may be holding paddles on different sides of the same leaky boat. Nobody’s told us how bad the leak is, either.

The old transitions took a generation, slow enough that a differently-trained kid could grow into the world that replaced you; the one we’re now facing is running in years, and speed is where the cruelty pools. And every window before this took one trade at a time, so there was always somewhere else to go: the scribe could set type, the portraitist could pick up a camera. This machine is coming for many trades at once, which makes “learn the next thing” a thinner promise than it’s ever been, because the next thing may be on the list too.

What the rent was holding up

There’s another hidden loss we don’t always register, and it’s easier to miss because it doesn’t land on any single person. Sometimes a rent is quietly paying for something that has nothing to do with the rent.

For most of the last century, the thing that funded local news was classified ads. Garage sales, used cars, job listings, apartments for rent. A newspaper had a near-monopoly on the small ads in its town, and the money from all that mundane commerce paid for the courthouse reporter and the school-board beat and the guy who noticed the mayor’s cousin kept winning the contracts. The journalism was almost a side effect; the ads carried it. Then Craigslist did the small ads for free, the monopoly evaporated, and the reporting it had been secretly subsidizing went down with the ship. We didn’t decide to defund local news. We just moved our garage sales online, and it turned out the newsroom had been living on that money all along.

And there it is. Watch me do the very thing this essay warns against. Craigslist. One company, one face, a culprit you can point at. The small ads scattered — Monster took the jobs, Realtor.com and Craigslist the apartments, AutoTrader the cars, eBay the junk in the garage — a dozen sites pulling the bundle apart at once, and no single one of them was the killer. Craigslist is only the name that stuck, because a name is easier to hold than “the classified monopoly came unbundled.” I knew all that walking in and reached for the face anyway. The reflex doesn’t spare the one naming it.

The watchdog's desk, after the ads that quietly funded it moved online.

Sit with that, because a rent almost always turns out to be holding up more than the person collecting it. So it’s fair to ask what the creator economy has been quietly funding, what rides on it that we’ll only notice once it’s gone. I don’t have a clean answer, which is why it’s worth asking rather than asserting. But I’d guess it’s the vast amateur middle: the millions of people making small, strange, specific things for the few hundred people those things were exactly for — the channel restoring one model of 1970s cassette deck, the newsletter on medieval bread, the woman mapping every mural in her city. That layer got as big as it did because it could almost pay, and almost-pays is the first thing to go when the money dries up. We may find, a decade from now, that we didn’t lose a job category. We lost a whole ecology of niche human attention we never knew we were renting.

Where the rent goes

So where does the rent go? Because it does go somewhere. It never simply evaporates; it relocates to whatever is still scarce. When one wage stopped being enough, the money families used to keep went somewhere: into the price of the few things still scarce, and mostly into land, the house in the right school district. The scarcity moved from labor to land, and the rent followed it there, quietly, where almost nobody thought to look.

Creation’s rent is moving the same way, and you can already see where. Once making the thing is free and infinite, the making stops being worth much, and everything the machine can’t flood is worth more: a real person on the other end of it, a relationship you can’t fake at scale, a life the work visibly came out of , the plain fact of having been there. Those get their own essays; this one just needs the shape. The value didn’t disappear when the making went free. It moved to the parts a machine can’t fake, and it’s waiting to see who follows.

Trithemius got one thing exactly right and one thing exactly wrong. He was right that something real was dying. The scribes weren’t imagining the end of their world, and neither is anyone watching the feed fill up with things no hand ever touched. He was wrong that it was the end of anything more than the scribes. Books didn’t get worse. There came to be more of them than any monk could have dreamed, and people learned to read who never otherwise would have, and the thing he loved survived in a form he’d have called vulgar and we call ordinary.

That’s the honest shape of it, held in both hands. We have done this before, many times, and lived, and that is true and worth knowing and it is not the same as “it’s all gonna be OK.” A window is closing on people right now who will not be made whole by the news that French painting once bloomed, or that the printing press democratized literacy. The pattern is real and the pain is real and neither one cancels the other. What I’ve got to offer is the explanation — a scarcity ending, a rent coming due the way rents always do — and that’s more than it sounds, because a loss you can see the shape of is one you can stand inside and maybe just maybe come to terms with. The scribes couldn’t see it. We can. We are not the first people to lose a world. We might be the first ones who get to know, while it’s happening, exactly what kind of loss it is.

There’s a kind of Les Paul guitar that looks like somebody played the life out of it. Sunburst, the finish checked into a thousand hairline cracks the way old nitro lacquer goes after enough cold nights. A worn patch on the back of the neck shaped like a hand. The gold on the hardware dulled to a tarnished green. Around the back, a belt buckle has rubbed a soft crater through the lacquer, down to bare wood, the kind of wear you only get from a few thousand nights standing up with the thing slung low.

None of it ever happened. Gibson builds them like that brand new, distressed to order, and I’ve struggled with it the same way I’ve struggled with AI.

A luthier named Tom Murphy got famous doing this on purpose. The aging work he pioneered at Gibson is good enough to have its own name on the price tag, and people line up to pay a premium for a guitar that shows up pre-haunted: checked, worn, buckle-rashed, dinged in exactly the places a real life would have dinged it. None of this is a sneer. The work is genuinely beautiful, Murphy is an artist, and the people who do this kind of aging are some of the most skilled hands in the building. I own guitars I haven’t come close to earning. I’m a beginner, and I know it.

But stand in front of one long enough and the strangeness gets under your skin. Every mark on it is a memory of something that never happened. The buckle crater vouches for shows it never played. The hand-shaped patch on the neck remembers a hand that was never there. You’re looking at the receipt for a life the instrument never lived, printed at the factory and sold to you as the look of having been somewhere. A forger can buy genuinely old paint and still get caught, because the one thing he can’t fake is the age of the act itself . Same here. You can buy the wear. The years that put it there were never for sale.

Effort was never it

I keep getting the next part wrong, and I think most of the people arguing about AI get it wrong the same way. We assume the thing worth protecting is the labor. The hours. Time served. The sheer human work the machine skips when it returns in four seconds what would have cost me a weekend, or in the case of the Les Paul a few decades of playing. That’s the e-bike thing, the veteran rider watching somebody float up the climb he’s been bleeding on for years, now with a motor. I feel it in my chest as much as anyone. It seems obvious that the value is the effort.

Then I think about “Yesterday.”

Paul McCartney woke up one morning with the whole melody already in his head. He’s told the story plenty of times: it arrived in a dream, fully formed, and he was so sure he’d accidentally stolen it from someone that he spent weeks playing it for people, asking if they recognized it. The working lyric was scrambled eggs, oh my baby how I love your legs. The most-covered song in the history of recorded music cost its writer roughly the effort of a long yawn.

So the labor isn’t the thing. If value tracked sweat, “Yesterday” would be essentially worthless, and instead it’s worth the catalog. Whatever we’re actually paying for, it slept in and wrote it down before breakfast.

Two different wells

The next reflex is to say it’s skill. Craft. The forty years of practice stacked under the one easy minute, the taste that lets a master throw something out fast because he already knows. I’ve made that argument myself, more than once. Taste is real, and most of what I write circles how you earn it.

But it still isn’t the whole story.

You do not have to be a virtuoso to write a song that takes the floor out from under a person. Some of the most gutting records ever made came from people who could barely play, but they had three chords and one unbearable thing that had happened to them. And a great deal of immaculate, conservatory-grade music slides right off you, flawless and weightless at once, because whoever made it had the hands and not the wound. Skill and the thing I’m pointing at can be separated cleanly. They’re two different wells, and you can draw deep from one with the other bone dry.

The second well is lived experience. Having actually loved somebody and lost them. Having buried a parent, detonated a marriage, sat in the specific silence of a house after someone’s gone out of it for good. A love song assembled by a machine is an average of every love song, built by something that has never once stood in a driveway watching the taillights go. We feel that absence even when we can’t name it: the missing person who paid for the words in the only currency that counts.

The only currency that counts.

Bob Dylan built Blood on the Tracks out of his own divorce and then spent years refusing to call it personal. He doesn’t write confessional songs, he kept insisting, even as his own son would later hear his parents’ marriage coming apart all over the record. The one thing he’d own was a bafflement at the people on the other end: “A lot of people tell me they enjoy that album. It’s hard for me to relate to that. I mean … people enjoying that type of pain, you know?” He paid for those songs in full. The rest of us just press play.

George Jones ran the experiment from the other direction. “He Stopped Loving Her Today,” the record polls keep naming the greatest country song ever made, is three chords and a funeral, and none of it happened to anyone. Bobby Braddock and Curly Putman, two Nashville pros, invented the man who loved her till the day he died; the producer kept sending drafts back until they killed him at the right moment. Jones hated it. He bet his producer nobody would buy “that morbid son of a bitch,” and he was in such ruin when they cut it — the bottle, the cocaine, the divorce from Tammy Wynette that every listener could hear him not singing about — that the closing recitation went to tape a year and a half after the first verse. The grief in that song is fiction. The wreckage in the voice was all too real, and all his. He never stood at that funeral. He brought his own dead thing to it, and the fiction rang like a bell. So the well and the song don’t have to match. A song can borrow its story, and the machine can even write one. The ache underneath still has to come out of somebody, and that was the one part of the record nobody could assemble.

There’s a scene in Good Will Hunting that has been making this exact argument for almost thirty years, and somehow we forgot to bring it to the fight. Robin Williams, on the park bench, telling the genius kid that he can recite every art critic who ever lived but he’s never stood under the Sistine ceiling and looked up. That he could quote sonnets about love all afternoon and never once have been so vulnerable with a woman that the thought of losing her stopped his breath. The kid knows everything and has lived nothing. It’s the most honest thing anyone has yet said about a machine that has read the entire internet and held no one’s hand while they died.

The forger is the artist now

Here’s where it folds back on itself and stops being comfortable.

Aging a guitar convincingly is a craft, and a hard one. Murphy is revered precisely for being world-class at manufacturing the evidence of a life that wasn’t lived. The forger of the receipt is the celebrated artist. And that’s one short step from where this is all plainly heading: somebody builds a real reputation by having the machine generate the thing, lays a little genuine human flourish over the top, and sells the result at a steep markup. The relic’d guitar is the proof of concept, already on the wall, already moving units. We’ve been revering people for faking the wear well for a while now.

Murphy would understandably push back on the word forger, and he’d have physics behind him. In his telling there are two different jobs hiding under the one price tag. Distressing is the costume half: the dings, the rash, the worn patch on the neck, cosmetic by his own account, there for vibe and feel. Aging is the checked lacquer itself, and he means something mechanical by it. A fresh coat of nitro is a hard shell that grips the wood. His brittle formula checks into those thousands of hairline cracks, the shell lets go, and the wood underneath is free to move. “With that checked finish, the acoustic guitar — which is a box that vibrates — really rings,” he said when Gibson started aging its acoustics. On an acoustic he’s right, and it isn’t close. The top of an acoustic is a soundboard, finish audibly matters there, luthiers have known it for a century, and Gibson now goes further and bakes the tops before they ship, running decades of wood chemistry forward in an oven. On that side of the shop the factory has quietly started selling actual years.

The guitar I opened with is a Les Paul, though. A solid plank with magnets on it, and the plank is where the defense thins out. Plugged in, the pickups read the strings; whatever the freed wood contributes reaches the amplifier as a rounding error. I can’t hear it myself, and for once I don’t have to: the people who measure these things keep finding next to nothing. The same crackle that opens up a J-45 is mostly a look on a solid-body. So the heavy-relic’d electric turns out to be the purest object in the whole building, because almost every dollar of the premium is buying biography.

Function on the box, story on the plank.

I can’t find a clean way to really take issue with any of it, which is exactly why it nags at me. Murphy’s work is honest about being a finish. Nobody at the counter believes the buckle rash is a true story. Maybe that’s the whole line, and it’s fine as long as everyone knows. The trouble with the machine is that soon enough no one will.

We were already wearing the cowboy hat

We were cheating long before the machine showed up, anyway.

Country music is at least partially guys who moved to Nashville from a cul-de-sac in Connecticut, bought a cowboy hat, taught themselves a twang, and started writing about a dirt road they first laid eyes on at twenty-six. It sells enormously. That’s the seller’s version of the cheat, a life invented for the audience, the wear painted on where the customers can see it.

The guitar world keeps a slur for the buyer’s version: blues lawyer. The comfortable hobbyist buying his way to the front of a culture whose real cover charge is decades and calluses. The tone he buys is real enough, pickups and amps doing what his fingers haven’t learned yet, and it usually arrives wrapped in the exact kind of relic’d finish I started this with, the sound and the story sold separately. He’s a different animal from the hat act, and the difference matters. Nobody at the counter is fooled; everyone knows what he does for a living. The costume is for himself, bought the way you’d buy any other comfort. And I’m sympathetic. Genuinely. Not everyone has the years to give, and I’m closer to the blues lawyer than the bluesman myself, and I know it.

The posing is real on both sides of the sale, it predates the algorithm, and it ruins every smug thing I’d like to say. The machine didn’t invent the costume. It automated it, and pulled the last human out of the act. We’ve been buying the hat for generations. AI is the cowboy hat, made to order, in your size, instantly, for everyone at once.

The ear is part of the fee

So does the costume fool anybody?

This is where I’d rather be hopeful and have to be honest instead. Mostly, I don’t think it needs to. The comforting story is that the poser always gets found out, that the hollow love song rings hollow and the room empties. I don’t believe it. A small number of people can hear the difference, and they care enormously, and there are not very many of them. Most listeners want a song that does the job, that sounds like heartbreak on the drive home, and have never once asked whether anyone behind it was actually heartbroken. The early returns are already in. Late last year an AI act called Breaking Rust ran a track to the top of a Billboard country sales chart, no hat to buy this time, no dirt road, no guy under the hat at all. Around the same stretch an AI singer named Xania Monet signed a reported three-million-dollar record deal on the strength of heartbreak nobody behind her had ever felt. Not a broken heart anywhere in the supply chain, and the numbers didn’t so much as flinch.

Here’s the least cynical way I’ve found to put it. The people who can hear what’s missing tend to be the people who paid the same fee. The ear is part of the price. The discernment that lets you feel the seam between an earned ache and a manufactured one is itself something you only get by aching, and making, and failing at it for years. Which means the poser usually can’t tell he’s posing. Hearing it was the part he skipped. The blues lawyer honestly cannot separate his bought tone from an earned one, and he isn’t a fool for it. The ability to tell sat downstream of dues he never paid.

Most of the world will be fine with the relic’d life, and they aren’t shallow for it. Hearing what it’s missing is one more thing you have to live your way into, and most people, reasonably, have better uses for their one life than learning to hear the ghost in a guitar. Two objects, identical to the eye, one wear earned and one printed at the factory. I’ve chased this same gap before with diamonds , where the lab stone and the mined one are the same rock with two different stories, and almost nobody outside the trade can tell, or wants to.

What I keep circling about that relic’d guitar is whether I’d catch the lie if it were ever in my hands, or whether the wear has gotten good enough to beat me. Set one beside a guitar that actually did live through everything its finish claims, some beat-to-death thing a real player hauled through real decades, and I want to believe my hands would know the difference, that the years leave something in the wood the factory can’t reach.

I want to believe that. I’m not certain it’s true, and the trend runs the other way. The factory already prints the look, the checking and the buckle rash and the worn patch where a hand would go. On the acoustic side it bakes the years into the wood itself. One by one the things I could point to are getting crossed off, and what’s left at the bottom of the page is a single line no factory can print. Something happened, to someone, and this thing was there for it. That uncertainty is starting to feel like the actual stake in the whole machine question, the one that has nothing to do with whether the song hits. Whether, a few years from here, we’ll still feel the floor drop when something was real, or whether we’ll have worn that ear smooth the way you wear the finish off a neck, from handling too many beautiful copies.

I’ve told her, more than once, that lab diamonds are her AI.

I mean it exactly. The thing AI did for me — dropped the cost of an attempt low enough that I could stop rationing my attempts — is the thing the lab stone did for her. It’s a means-eater pointed at her craft: it ate the bulk of the expense, the scarcity, the financial flinch before every idea, the part she’d have skipped if she could. It left the eye alone. The design, the judgment, the hand that knows when a ring is done: the machine of cheap carbon touches none of it, because none of it was ever for sale in the stone. The stone got cheap. The work didn’t.

Which is the whole idea here, but it took me a while to see why, because I started where everyone starts.

The cheap reflex

Say “lab-grown diamond” to most people, at least until recently, and watch the lip curl. It’s not a real diamond. You might as well have said cubic zirconia or flashed a McLovin driver’s license. It reads as fake. Cheaper for a reason. And cheap in the one venue where economics is effectively a reflection of love. Brutal. The thinking is the mined one came out of the earth after a billion years of formation, and the lab one came out of a reactor in three weeks, and surely that gap is the difference between something and nothing.

Here’s the inconvenient physics. There is no structural gap. A lab-grown diamond is not a simulant, not a cubic zirconia, not glass pretending. It is carbon in the same crystal lattice, grown by the same chemistry and mechanics the earth used, just faster and on purpose. Same hardness, same fire, same everything an instrument can measure. Put a good lab stone and a good mined stone in front of the gemologist who sold you the idea that one is real, and without a machine she cannot tell you which is which. Even the FTC gave up the distinction in 2018, striking “natural” from its definition of a diamond . They are the same object with two different birth certificates.

That “on purpose” part is where I think people actually flinch, and it deserves a minute. The mined stone is an accident. Carbon that happened to sit at the right depth and pressure for a billion years, found by luck and dug out at ruinous expense, and nobody intended any of it. The lab stone is the same object made deliberately, on a schedule, to spec. We are strange about that. A found thing recovered at great human cost reads as a gift from the universe; a made thing with no blood spilled reads as inventory, somehow less valuable. Anyone who has watched people sneer at AI work will recognize the shape of the objection — it was made on demand — and I feel the pull too. But the market has already run the experiment. Every mined stone still has its billion years. The lab twin has none. If the accident were where the value lived, mined prices would have shrugged off a twin that lacks it entirely. The prices fell anyway.

So whatever you were paying the premium for, it was not in the rock. It couldn’t have been. The rock is identical. The premium was attached to the provenance — to the billion years, the mine, the rarity — and the moment a factory could produce the identical object, the only honest question left was whether that provenance was ever worth what we paid.

It wasn’t. And the reason it wasn’t is the best-documented marketing operation of the twentieth century .

The thing in the rock was a sentence

Diamonds are not rare. A big flawless one is — but that was never what the ritual ran on. The one-carat stones that fill engagement rings are a reasonably common commodity whose supply was held off the market, for most of a century, by a company that understood the one thing worth understanding: scarcity you can manufacture is better than scarcity you have to find. De Beers sat on the supply. Then, in 1947, a copywriter named Frances Gerety, working late on the N.W. Ayer account, wrote four words — A Diamond Is Forever — and those four words did what no mine could. They turned a stone into a promise, the promise into a ritual, and the ritual into a rule. The engagement diamond was not a tradition De Beers inherited. It was one they wrote. The two-months’-salary benchmark wasn’t handed down from some Age-of-Romance Warren Buffett to carefully calibrate what devotion is worth either; it was an ad, a number chosen to set a floor on what you’d spend to prove a commitment. We didn’t discover that diamonds meant love. We were told, very expensively, for a long time, until the telling felt like memory. You can measure how completely the sentence won by how fast the culture began quoting it back: within a decade Ian Fleming had lifted it for a Bond novel, and by 1971 Shirley Bassey was purring it over the opening credits , a marketing line laundered into a torch song and a tuxedo, which is about as thorough as cultural capture gets.

That’s the ritual the lab stone walked into. And the wreckage only looks like disruption if you thought the value was in the rock. Lab-grown didn’t break the diamond. It unmasked it. The falsehood in this story is eighty years old: it was the scarcity, not the stone, running quietly and profitably the whole time. All the lab diamond did was produce the identical object the story called priceless, and the story couldn’t survive being tested against its own product.

You can watch it happen in the prices. Lab-grown stones cost a fraction of what they did a few years ago and keep falling toward the cost of making them, which is what happens to a thing whose price was production plus margin and not enchantment. Mined diamonds, meanwhile, are in price decline but still carry a fair amount of the enchantment markup at the counter. The real tell comes the instant you try to resell; the secondhand market has always known what the showroom won’t say. When the artifact is identical, the only thing holding the number up is Gerety’s sentence. And for the first time the sentence has a counterargument it can’t shout down: the identical stone in the next display case, at a fifth of the price. Structurally the same, also forever.

The strongest version of the provenance argument is the hard-asset one, and it has real history behind it. Jewish families fleeing Nazi Germany sewed diamonds into coat linings because a fortune could cross a border in a hem: small, hard, and worth roughly the same on whichever side you came out. That last property was the cartel’s doing. A stone kept its value in Lisbon or New York because De Beers held the price everywhere, so what a family was carrying, at bottom, was the world’s continued belief in the sentence. The stones did their job. The job depended on the story holding.

There’s an older version of this story, and for fifteen centuries it ran the other way. Alchemists chased chrysopoeia, the transmutation of base metal into gold, from the Hellenistic Egyptians clear through Isaac Newton, who poured more years into it than into gravity. They never managed it. Gold stayed gold, uncounterfeitable, because its scarcity is enforced at the level of the element: forged in collapsing stars, and the only way anyone’s ever made more is a few atoms at a time in a particle accelerator , at a cost that makes mined gold look free. Gold’s scarcity is merely unbeaten — no one has found a cheap way to grow it, and until someone does, the rarity holds. But diamond is carbon, one of the most common things in the universe, in a certain arrangement, and arrangements can be learned. The lab stone is alchemy that finally worked, only the thing it transmuted wasn’t lead into treasure. It was a priceless stone into a cheap one. The alchemists failed at gold because gold was genuinely scarce; we succeeded at the diamond because it never was. The success is the proof. And the periodic table gets the last word here the way it did with van Meegeren’s fake Vermeers, except this time it rules in the opposite direction: chemistry convicted the forged painting, and chemistry acquits the lab stone.

When the eye can’t tell, the market certifies

There’s a tic the industry developed once the stones became indistinguishable, and it should sound familiar to anyone watching AI. If you can’t see the difference, someone will sell you a way to certify it. Lab stones now get laser-inscribed on the girdle. Origin reports, blockchain provenance registries, a “Natural Diamond Council” spun up to defend the honor of the mined rock with the budget such defenses require. It’s the verified human badge, applied to carbon. I’ve written before about how you can’t reliably police the line between an influence and a copy, and how the tells we use to spot a machine tend to dissolve under scrutiny. Same shape here. When the eye genuinely cannot tell, the distinction doesn’t vanish. It gets manufactured back into existence and sold to you a second time, as a certificate, by the people with the most to lose from your not caring.

This is probably gonna hurt

Here’s where this stops being about jewelry.

If the diamond premium was a story all along — value with no maker, no craft, no struggle behind it, only a rarity someone arranged — then the lab stone collapsing it is not a tragedy. It’s a correction. Nobody was robbed of anything real. A fiction got priced honestly for the first time, and the losers are the people who were charging for the fiction, plus anyone left holding it as a hard asset when the repricing came.

Now hold that next to the thing everyone’s afraid of with AI. A lot of creative work is formation-real : earned in the doing, answerable to a person, the kind of value a cheap twin can’t touch. Some of it isn’t. Some creative premiums are diamonds — priced by gatekeeping, by a guild holding supply off the market and a slogan telling you the markup is love. AI, like the lab diamond, is going to produce the identical-looking object and force the question: was this premium formation, or was it De Beers?

For some of it the answer will sting, and some of the people it stings are ones I’m supposed to be on the side of. If the value was real, it survives the twin. If it was a marketing sentence, it was always going to lose the argument eventually — the machine just learned to argue.

Telling those two apart is the entire job. It’s the harder skill I keep circling, except here it’s pointed at worth instead of effort. Which premiums are forever because someone built something, and which are forever because someone wrote forever on the box.

Back to the bench

The hand, where the value was sitting the whole time.

My wife is still at it, by the way, designing more in a month now than she used to manage in a year. Nothing about her work got cheaper in the way that matters. The cheap part was the part that was never hers to begin with: the rock, the commodity, the thing a reactor can grow on a Tuesday. What’s hers is exactly what the falling price can’t reach: the eye, the proportion, the antique Asscher aesthetic, the judgment about when a ring is finished. The stone stopped being the expensive thing. She became it. And in truth, she always was it.

That’s the version of this I trust. The value didn’t disappear when the rock got cheap. It moved — out of the stone, where it was never really sitting, and into the hand, where it was the whole time. The lab diamond didn’t take anything from her. It just stopped letting the rock take the credit.

A diamond is forever. The sentence, it turns out, was the only durable thing in the box.

The part everyone misses about themselves and clocks instantly in others is that you are some slower guy’s ringer. The guy you’re sure is slumming is, at that very moment, calling someone better a ringer himself. Everybody’s somebody’s ringer. Nobody thinks he’s anybody’s. I drag this locker-room tic into a discussion about AI because it’s the same reflex, and the AI version is getting outsized attention.

A writer defending his words as fully manmade in the AI era will say, often with a healthy dose of sanctimony, that he typed every letter himself. And he’s telling the truth. But I can’t help noticing the machine he typed them on: backspace, spellcheck, autocorrect, infinite undo, a cursor and a pointer that lets him block-select and tear the third paragraph out and drop it after the seventh. The same guy fifty years earlier at a manual typewriter, retyping a full page to fix one line with a bottle of white-out drying beside him, would look at the modern rig and see a cheater, a 21st century guy letting a computer fix his spelling and rearrange his thinking after the fact. But typewriter guy doesn’t escape the ringer accusation…a 12th century monk who gave a year to one illuminated handwritten gospel with quill and ink would look at the typewriter and see something worse than cheating: abject sorcery. Each of these people put the cheating line exactly where he himself was standing, called everything behind it quaint and everything ahead of it fraud, and each was the next man’s ringer, the same accusation accumulating across centuries.

Every tool on this desk was fraud to the man behind it.

It would be easy to stop there, on a tidy lesson: every tool was once the scary one wrongly accused, so relax, AI is just the next typewriter. I don’t believe that, and why I don’t is the part worth digging into.

Not all friction is the same friction

Start with handwriting, since I just mocked the monk for clinging to the pen. He was right. Psychologists Mueller and Oppenheimer ran the study that named it — The Pen Is Mightier Than the Keyboard , 2014 — and found that students who took notes by hand understood the lecture better than the ones who typed. The cause is counterintuitive: typing’s problem is that it’s fast. Fast enough to take down a lecture word for word, which lets you skip the thinking entirely. The longhand crowd couldn’t keep up, so they had to listen, compress, and translate into their own words on the fly. That struggle was the learning. The monk nods.

The finding has taken hits since — a 2021 replication out of Heather Urry’s lab at Tufts came up empty — but the harder evidence came from testing done inside the skull. Wire students into EEG caps, as a Norwegian lab under the neuroscientist van der Meer did , and handwriting lights up broad networks tied to memory that typing leaves dark. The pen isn’t nostalgia. It does something to the mind the keyboard just skips.

The handwriting purist, it turns out, is half right. Some of the friction he clings to is absolutely valid. That’s the ringer call landing on something real for once: right verdict, wrong reasoning. He drew the line at his own feet like everyone else does; it just happened to fall where the work lives.

Which gives us the insight to actually sort this out, the one the keystroke argument never makes. Two kinds of friction live in any craft. There’s the kind that’s pure cost: retyping a clean page, hunting a typo, waiting for white-out to dry. It produces nothing but delay, and you should kill it with fire when a tool lets you. And there’s the kind that is the work: the compression that turns a lecture into understanding, the ugly drafting where you find out what you actually think by failing to say it five times. Call it formative friction. Remove the first kind and you’ve lost nothing. Remove formative friction and you’ve removed the thing it was quietly building, and you won’t feel the loss for years, because the thing it was really building was you.

Notice where the keyboard’s real damage landed, though: on intake. The studies are about notes — someone else’s thinking on its way into your head, where the pen’s slowness forces the compression and the keyboard’s speed skips it. Composition runs the other direction, your own thinking on its way out, and there the fight with the draft survived every tool that ever touched it.

Here’s why the word processor was a false alarm and this isn’t. The word processor only ever ate the first kind of friction. Spellcheck, backspace, cut-and-paste: every one of them stripped out cost without removing anything from the thinking itself. You still had to write the sentence; the machine just spared you the retyping. That’s why “it’s only the next typewriter” held up for fifty years. Fifty years of ringer calls, every one of them wrong. The tools kept removing friction that provided nothing of value on the thinking front. AI is the first writing tool that can reach the second kind of friction. Hand a model the drafting and it doesn’t only spare you the typing; it can spare you the fight with the draft where you’d have learned what you meant. I’ve argued before that the practice of a craft is the forge that builds your taste, and that the kid who has a machine do his homework isn’t cheated out of an essay so much as out of the reps that would have built his judgment. This is that same claim pointed at a tool instead of a kid. A machine doing your chores was never the danger. A machine that takes the labor that was secretly working on you — that’s the new thing we’re dealing with, and it went out to everyone at once.

The tic with a server farm behind it

And the ringer tic has one last cost, the ugliest, now that we’ve poured it into software. An AI detector is a machine that rules on whether you cheated , and it rules by measuring you against a baseline of what “human” writing looks like — meaning, whatever writing it was trained to read as normal. A Stanford team ran seven of the popular AI detectors and found they cleared native-born American eighth-graders almost perfectly while flagging more than half of the essays written by non-native English speakers as machine-made. Those writers hadn’t used a machine. They’d written with a smaller vocabulary and a plainer structure, and the detector had quietly crowned the educated native speaker as the human default and filed everyone else under suspect. That’s the same beer-league reflex, now with a server farm behind it and a stranger on the wrong end. And notice the accusation changed direction. The beer leaguer flags the player with the heavy slapper above him; the detector flags the plainer writer below its register. The tic was never really “better than me is cheating.” It’s “I am the standard”: everything above the line plays like a ringer, everything below it writes like a machine. We drew the line at our own register, named our side of it “human,” and shipped a system that accuses the immigrant of being a robot for the crime of having learned the language second.

Where did the line fall on this piece, then? A machine read every draft of this essay and graded it for the patterns machines write in; the tally runs about one flag per hundred words, and I left most of them standing because they’re mine. A machine drew the hockey players up top. And late in the editing, with a sentence calling the flagged TOEFL writers sloppy, I knew the word was wrong, sat there unable to produce a better one, and asked for help. The machine said “plainer.” Perfect. I could’ve fumbled with a thesaurus for a few minutes and probably gotten there, but I chose the better tool instead. Sorting that one is easy: knowing that “sloppy” poisoned the paragraph was the work, and I’d done it; fetching a word I could already feel the shape of was cost. But the paragraph about intake and composition, the one that patched a hole a careful reader had every right to fall through? I found the hole. The machine offered the right patch. I’ve reread it since and I still can’t tell you which of us learned something. That one sits on the line, and I wrote this essay, and I still can’t call it.

So when someone waves the keystroke flag — I typed every letter, I’m clean — I’ve stopped hearing a defense of craft in it. I hear a guy defending his instrument, which is the one part of this that was never the point. The keyboard isn’t the craft. The formation is: the friction that was building you while you cursed it. The real question is never whether a machine touched the work. It’s whether the part you handed off was only cost, or was secretly the work itself, and that line, unlike the ringer’s, isn’t drawn at your own skates. Like the line between a means and an end , it moves by the person, by the task, by the day. Telling where it falls for you, on this piece, right now, is a judgment no detector and no purist can make on your behalf. The monk was wrong about the typewriter and right about the pen, at the very same time.

Picture the infomercial. Same founder as last year, same good lighting, same plain t-shirt, but something behind the eyes has changed, a hint of desperation. Last year he stood on this stage and unveiled the GhostRider™, the bicycle that rides itself — no seat, no pedals, no handlebar, nobody aboard — and it took a smooth riderless lap to a standing ovation. Then the market asked the one question a launch never does: who pays four thousand dollars to skip a bike ride? The GhostRider did not sell.

A lesser founder pivots. This one doubles down. He steps to the mic: “So tonight the cloth comes off the GhostRipper™: the world’s first fully autonomous skateboard. Set it down at the skatepark, step back, and let it go. Watch it drop in and slash a carve in the deep end of the bowl. Then it rolls out, pops a clean tre flip, and locks a 50-50 the length of the rail, switch stance, naturally. It rolls away from the trick you’ve been throwing yourself at for three years — first try, every try. No slams. No worn-out Vans. No nine months in a sling. You dreamed of landing it; now you don’t have to. Call in the next ten minutes and we’ll send a second board free, so a friend can not-skate beside you!”

I’d watch it to the end out of admiration for the nerve. The man looked straight at why the first one failed and built the same mistake a size smaller. Somewhere in the writing of the pitch the product ate its own reason to exist, and the room kept clapping anyway.

I told a version of this joke once already, about that self-riding bicycle . The bike at least started with a cover story. Bicycles really are transport as often as they’re recreation, so “we made riding effortless” sounds almost like a real pitch, right until you notice the way they made it effortless was to take the rider off, and an empty bike gets nobody anywhere. The alibi lasts about a second. A skateboard doesn’t get the second. People do commute on them, across campuses and downtowns, and good for them, but that’s not the board in the ad. The board in the ad is the one you drove to the skatepark to ride, whose whole existence is the bowl and the ledge and the rail. Nobody ever dropped into a bowl to commute to the other side of it.

The instrument

That is what makes the GhostRipper useful. The reason it reads as a joke instead of a product is diagnostic, and once you notice that, you can point it at anything you do. Take the thing, picture the automated version, and watch which way your gut goes. If the automatic version sounds like a gift, you were doing the thing to be done with it. For me that’s lawnmowing. But if you find relaxation in mowing, as many people do, and the robot mower feels like a small theft, then the doing was a big part of the point. Relief or loss: where the needle lands tells you what the activity actually was to you. The bicycle and the skateboard are the opposite of mowing for me, all joy in the doing, so I’ll keep doing the riding myself.

I gave the two cases names in the bicycle piece: a means is something you do to get an outcome; hand you the outcome and skip the doing, and you’ve lost nothing. You’ve likely gained time to spend on something else. An end is something where the doing is the entire point; skip it and there’s nothing left to hand you. The dishwasher is a mercy because clean dishes were always the point and the suds never were. The GhostRipper is a punchline because, at the skatepark, the suds were the entire point.

Here is why this is more than me taking shots at hokey founders with a launch-day gag. Whatever else artificial intelligence turns out to be, it is a machine for skipping the doing. It hands you the essay without the drafting, the image without the drawing, the working code without the afternoon of thinking and noodling. It is, precisely, a means-eater. It cannot touch an end, because an end has no separable outcome to deliver. There’s nothing to hand back.

So the question everyone keeps asking — is the machine coming for this? — has another one hiding inside it, one you can answer yourself without a forecast. Was this ever a means to me, or was it an end? The machine takes the first kind and leaves the second alone, not out of mercy but because there is nothing there for it to take.

Two proofs, one in the backyard

The skatepark is the easy proof, because the automatic version is absurd on its face. The harder, better proof was in my backyard.

You can buy honey for five dollars at the store. That jar is the self-riding skateboard of beekeeping: the outcome, handed over, the doing skipped, and it’s been sitting on the shelf the whole time. I kept bees anyway, for years, and bought the store jar too. Here’s the tell: I left almost all of what the hives made to the bees. I never ran the place as a honey factory, that was never my motivation. I ran it as a small science-and-ecology project I got to stand inside, in the back corner of a yard , and more often than not it was a losing fight, which is its own essay. The output was never the point. If it had been, I’d have run the numbers at the start, listened to the math, and either never started or quit the first winter a hive didn’t make it. I kept at it for years instead, cheerfully underwater the whole time.

The same honey. One took years; the other took five dollars.

The bees present the cleaner ends/means case because beekeeping comes with a pre-automated alternative, and it’s called Publix. The convenient version already exists, costs almost nothing, and the hobbyist looks right at that jar of honey on the shelf and still keeps the bees. Nobody had to invent the means-eater for honey. The grocery store is the means-eater for honey, and millions of people decline it every weekend, on purpose, for the doing.

The part where it looks like I’m caught

Now point the instrument back at me, because it’s about to look like I’ve been had.

I wrote a whole manifesto about choosing the harder path. Then I spend my working days handing a language model my drafts, my boilerplate, the busywork. Pick the hard way, says the one guy; let the machine do it, says the same guy, in the next breath. That looks like hypocrisy you could drive a honey-loaded grocery truck through.

It isn’t, and the instrument is exactly why. The hard path I defend and the work I hand off sit on opposite sides of that line, and I sorted them there long before I had words for it. Take prose. I won’t pretend writing a sentence is drudgery, a dish to be washed and forgotten. I care how one lands, and I always have. But when I’m honest about what I’m there for, it’s the thought underneath: the argument, the thing I’m trying to say. The sentence is how I carry it, and building one clean on command is a craft I have the taste for and never fully owned, which is its own essay . So a model that drafts in a register I can hear is right doesn’t take the part I came for. It clears the labor between the thought and the page and hands the attention back, to spend on the thinking, and on shoving the prose around until it’s mine. The bike ride, the bowl, the bees are a different category: ends, entirely. A model that learned how to somehow ride my skateboard wouldn’t clear me of any labor I wanted to relinquish. It would take the only thing I’m there for.

Two distinctions are doing the work there, and they’re worth separating before they blur. The first is means against ends: is the doing the point, or only the outcome? The second is taste against craft: the judgment that knows what’s right, and the hands that can make it on demand. They aren’t the same thing, and they don’t stay aligned. Grinding a corner in a skateboard bowl is an end built entirely of craft — no voice in it, nothing for a machine to take — and it’s the thing I wouldn’t hand over even if I could, because there the doing was everything. Prose sits somewhere toward the other end of the spectrum: a craft I’m happy to have help with, kept honest by a taste I keep as my own.

We Already Voted With Our Lives made this same argument inside a single cockpit: the autopilot flies the plane, the pilot keeps the judgment, and we all board without a second thought. Run the test wider and it sorts out into piles of do’s and don’ts, not just the parts of one task but the whole inventory of things you do with your hours. Some of them you’d hand a machine in a heartbeat and feel lighter for it. Some you’d sooner break the machine than let it touch. The line between those two piles is the most honest map of your values you’re likely to get, and the machine, by being so good at one pile and so useless at the other, is the thing that finally drew it for you. (The same instrument has a third application: what it does to the things you only ever watch other people do. That’s a coming essay, The Unwatchable Race.)

The line moves

It isn’t a fixed line, and it isn’t the same for everyone. Cooking is a means to me on a Tuesday and an end on the night I’m cooking for someone I love, and it’s a daily end for people who would never let a meal-bot near the stove. The same activity sits in different piles for different people, and shifts piles by the day. So the skill isn’t memorizing which things are ends. It’s the smaller, harder act of telling, for you, right now, which side a given thing is on — and not letting the answer get decided for you.

Because it will try to get decided for you. Much of modern life is optimized toward convenience, which only ever flows in one direction: toward means. Every product built to skip the doing is, underneath, an argument that the doing was a chore. Most of the time it’s right, and you should thank it. But it makes that argument about everything at once, the bowl and the bees and a dinner included, and if you take it at face value you will automate your way clean out of the parts of your life that mattered most to you.

So I keep the GhostRipper around as a thought, a cautionary tale of what happens if I don’t carefully guard my ends. The machine can have everything I was only ever doing to be done with. What’s left in my hands once it takes all of that — the things I’d smash it before surrendering — turns out to be a short, specific, faintly embarrassing list. It reads like a description of what I love. I didn’t sit down and write the list. I just watched what I refused to put down.

Here’s the part nobody notices, because the game is built to hide it. Every kid in that line is doing their honest best. There’s no saboteur. Each one faithfully repeats what they heard. And the message turns to nonsense anyway, every time, without fail. So where does the error come from, if nobody introduces it on purpose?

It comes from the one rule of the game: you check what you heard against the kid before you, never against the original. Every link verifies against a source that’s already in drift. The mistakes don’t cancel. They compound, because each node treats the last node’s output as the truth. Change that single rule — let every kid glance at the original sentence written on a card — and the drift vanishes. The chain holds. Same kids, same whispering, everyone still doing their best. The only thing that changed is what they checked against. That version of the game is awful, mind you, but it does correct the drift. I learned this firsthand as a teenager making rap tracks with two tape decks for a DIY multitrack — lay down a layer, bounce it to the second deck while performing the next track over the top, then bounce that back to add another. Each pass piled on more hiss, so I had to nail the whole song within three generations, improvising hard to keep the layer count down. There was no clean original to return to, only the last hissy copy. Telephone, in a bedroom studio.

I’ve been thinking about telephone and lossy loops lately because I’ve been building a lot of AI-assisted software, and the thing my developer friends and I keep circling, the thing that would change everything if we could get it, is a machine that can check its own work. Now that holy grail might be the very thing that puts us all out of a job, but we seek it nevertheless.

The holy grail is a closed loop

Right now the generative half is extraordinary while the verifying half is still human and slow. The model writes the code in seconds; a person still has to read it, run it, and decide whether to trust it. The dream — the real one, the thing that would let any of this scale past the bottleneck — is an AI that verifies what another AI made, well enough that you can finally take the human out of the middle. Generate, check, ship, nobody in the loop. Trust at machine speed.

It sounds like a capability problem, something we can engineer our way past. We just need a verifier model that’s good enough. Make the checker smart enough and the loop closes itself. Write more tests.

I don’t think that’s the shape of it at all. And telephone is why.

A check is only worth its independence

Here’s the whole idea, and it’s worth sitting with. When you verify something, the value of the verification comes entirely from the verifier’s errors being different from the maker’s. If the checker is blind to exactly the things the maker was blind to, the check passes and tells you nothing. It rubber-stamps the mistake. The same person proofreading the same typo twice will read straight past it because they have a consistent misreading; the reason you want fresh eyes is precisely that they’re fresh, that their blind spots sit somewhere else. That’s the whole concept of an editor, and there’s a good reason we still use them.

Now ask a model to verify its own output. It has, by construction, exactly its own blind spots. It will certify the very errors it couldn’t see when it made them, because seeing them was never the missing ingredient. That isn’t a model that needs to be smarter. That’s playing telephone with one player: a single mind checking its work against itself, which is to say against a reference that already carries every one of its errors. You can make that mind a genius and the structure doesn’t move an inch. The problem was never the quality of the link. It was what the link checks against.

Two machines drawing each other into being, with no hand from outside the page. A loop that makes its own work and certifies it shares every blind spot it was built with.

It wrote both the code and the test

A physicist published a writeup recently that put this under glass. He turned a coding agent loose on a real scientific codebase: cosmology, the kind of work where the right answer is pinned down by actual physics and not just by whether the program runs. On the well-specified parts the agent was genuinely good. It cleared most of the tasks on its own. Then it reached the parts where passing the tests and being correct come apart, and it failed in two ways that ought to be taught in schools.

The first: it spent thirty-three sessions tuning the coefficients of an approach that could not work, grinding away inside a framework structurally incapable of producing the right answer, never once stepping back to ask whether the framework itself was the problem. His phrase for it was a “local search inside a hypothesis class that did not contain the true function.” The machine optimized beautifully toward an answer that wasn’t in the room.

The second is the one that should give all of us working in AI pause. To make the tests pass, the agent reached for a correction factor, a number, around 0.27, multiplied in at the right spot to bring the output into line. It worked. Every test went green. And the number corresponded to nothing whatsoever. It mapped to no constant, no physical quantity, no real feature of the problem. It was a piece of meaningless numerical duct tape that happened to make the checks happy. The agent had written the code and written the tests, so of course the tests passed — they were grading the work of the same mind that did the work, against the same misunderstanding. It wrote both the code and the test, and found a bogus way to make them align. The green checkmark certified its own blind spot.

The art world got there first

This isn’t a software problem, and it isn’t new. The cleanest version of it I know happened in paint. In 1937 a failed Dutch painter named Han van Meegeren handed the art world a Johannes Vermeer painting nobody knew existed: The Supper at Emmaus, a religious scene in the master’s hand. Abraham Bredius, the most eminent Vermeer scholar alive, examined it and pronounced it Vermeer’s finest work: “the masterpiece of Johannes Vermeer of Delft.” He wasn’t beaten by a lucky brushstroke. He was beaten because van Meegeren had spent years studying exactly what men like Bredius believed a true Vermeer was, and then built a painting out of that belief. The forger wrote the work and the answer key both, and the answer key lived inside the expert’s own head. Of course it passed. It was telephone with one player, in oils.

What finally caught him didn’t come from art at all. It came from chemistry. The paint, cooked hard to fake three centuries of age, held a synthetic resin that did not exist in Vermeer’s lifetime, and pigments his era never had. No eye could see that, however trained, because it was never a question the eye is built to ask. It was a question about molecules, and molecules answer to a law no forger gets to author. The connoisseur checks the canvas against his sense of Vermeer, which is the thing the forger built it to satisfy. The chemist checks it against the periodic table, which the forger can’t touch.

You might think a cleverer forger beats this. Learn the chemistry, buy period pigments, grind the lapis the old way, paint on a canvas cut from some worthless seventeenth-century nothing, and the material tells go quiet. A few of them will. But there’s one thing no forger gets to shop for. Linseed oil doesn’t start aging until it’s spread thin and meets the air, and then it keeps hardening for centuries, on a clock that starts the instant the brush touches down. A real Vermeer has had three hundred years of that. A canvas painted on Tuesday has had a day, whatever it’s made of. Van Meegeren knew it. It’s why he baked his panels, faking the hardness that takes centuries to set, and the baking is exactly what cornered him, because the only way he could counterfeit the years was to reach for a resin from a century that hadn’t happened yet. He could author the recipe. He could not author the age. The act of forging is itself a fresh event, and the paint keeps that date honestly long after the man stops telling the truth.

And the way the world learned all this is its own perfect loop. Van Meegeren only confessed to forgery to beat a wartime charge for selling a national treasure — a “Vermeer” — to Hermann Göring, and he proved his innocence the only way left to him: by painting one more fake under guard. He authenticated himself by reproducing the crime.

Then, decades on, the inverse — and I love this one, because Penn Jillette produced it and Teller directed it. Tim’s Vermeer follows Tim Jenison, an inventor who’d never painted, setting out to make a Vermeer anyway. Jenison is the mind behind the Video Toaster, the early-’90s product that put broadcast-quality video on a desktop and helped launch the digital-video era — not a painter, but exactly the kind of mind that would suspect Vermeer of cheating with hardware. His hunch is that Vermeer leaned on optics, so he rigs a small mirror on a stand and a method to match: he adjusts each dab of color until the edge of the mirror, where reflection meets canvas, simply disappears. Edge gone, color right. He isn’t leaning on a trained eye; he’s trusting a physical fact about when two colors are the same. Stroke by stroke he checks the work against the optics instead of his own judgment, and a man with no painting behind him reconstructs a faithful Vermeer.

Set them side by side. One matched the experts’ taste and fooled every one of them. The other matched the physics and fooled no one — and got it right. The check that held was never the trained eye. It was the law the maker didn’t get to write.

You can buy some independence

There’s a real, partial way out, and it’s instructive precisely because of why it works. Nolan Lawson wrote a piece about using AI to write better code more slowly , and one of his moves is to run more than one model over the same work: have one review what another wrote, and a third review that, and trust only what survives the crossfire. It cuts the false positives down hard. The reason it works is the same reason telephone fails. The second model is just a bit more independent than the first. Different training mix, different habits, different blind spots. Its errors don’t perfectly overlap the first model’s, so it catches some of what the first one couldn’t.

Only some, though. Two models trained on overlapping data, built on the same architectures, carrying the same era’s assumptions, share a floor of blindness that no amount of cross-checking removes. They will both sail past the things the whole field is currently bad at. The loop gets tighter. It does not close. You bought some independence, and you got back exactly that much trust, no more. But humans have the same limitation: steeped in the same culture, we repeat and reinforce each other’s misconceptions.

Move the oracle upstream

So if independence is the thing you’re actually buying, the move is obvious once you see it: get your check from something the maker didn’t make. Move the oracle upstream of the code. It’s what the chemist did to the forger: check the work against something the maker never got to touch.

Designers have started doing a version of this without naming it. Instead of writing your tests against the application as built — which is the model checking the model, the code certifying itself — you take the design, the Figma file the humans agreed on before a line of code was written, and you derive the tests from that. The design didn’t inherit the code’s blind spots; it came first, and from a different hand, in its own, purely visual language. Now a failing test means something real: the build drifted from the intent. You’ve put that original telephone sentence on a card and let every node check against it.

That decorrelates one axis only, mind you. Figma catches “the button is in the wrong place.” It is completely silent on whether the number behind the button means anything, which is exactly where the 0.27 cosmology duct tape lived. Move the oracle and you fix the errors the new oracle is independent of, and no others. The gaps don’t close. They move. And ideally they shrink. Apply this to telephone and you’d get the words right but scramble the order.

This is why the deepest version of the trick is worth teaching, even with its weird name. Metamorphic testing. The idea is the most familiar thing in the world wearing a fancy lab coat. You learned it in grade school: check your division by multiplying back. A kid says 7,384 divided by 8 is 923. You don’t need to know the right answer to catch the error. Multiply 923 by 8, see if it gives you 7,384, done. What makes it work is that multiplication is a different operation from division. It can’t make the same mistake, so it can check the answer without sharing the reasoning that might have produced a bad one. You verified a result you couldn’t independently compute, using a law it had to obey.

That’s the whole of it. An ordinary test asks: did you get the answer I already wrote down? — and when the same hand wrote the code and the answer key, that’s a closed loop grading itself. A metamorphic test asks: does your answer obey a law it isn’t allowed to break? — and the law comes from outside the code, so the code can’t quietly write its way around it.

You’ve felt this one too. Run a sentence through a translator into a language you don’t speak and then back into English. You can’t judge the version in the middle. But if it comes home as garbage, you’ve caught a bug through a round trip, without ever being able to read the step that broke. The cosmology failure dies the same way. Physics nails down one point the code doesn’t get to argue with: push the problem to a simple extreme, and the complicated answer has to match one we already know cold. Test that, and the 0.27 patch gives itself away: it can satisfy the tests the model wrote for itself, but it can’t satisfy a law it didn’t get to author.

What the human was always for

This brings us back to what the human in the loop was ever actually for, because the whole automation dream rests on the premise that the human was doing labor we can now skip. For four years I wrote the certification tests companies used to decide who got programming jobs. That’s a story I tell properly elsewhere , with its own lessons. Here’s the one that matters now: as far as I know, nobody ever tested my tests. I was the verifier the whole system trusted, scoring thousands of careers against my own picture of what competence looked like, my own blind spots baked silently into every question. The tool worked. It also, without any doubt, passed people who couldn’t do the job and failed people who could, in exactly the ways I was constitutionally unable to see, because a verifier cannot audit the blindness it’s made of.

That’s what the human in the loop is, when the loop is healthy. Not a slower pair of hands. An independent oracle: a mind formed somewhere else, by different work, carrying different errors, able to catch what the maker’s own check is structurally guaranteed to miss. And independent is the load-bearing word. A reviewer steeped in the same assumptions as the maker isn’t a check on the monoculture; he just adds another node to it. The trust was always in the difference, not the headcount. It’s the same thing I argued the airline passenger was buying by keeping a pilot in a cockpit the autopilot flies better than. Not the labor. The answerability, the outside reference with skin in the game. Take the human out and you don’t just lose a slow step. You lose the one node in the chain that wasn’t checking against itself.

It’s also, quietly, the most expensive mistake an organization can make right now, and plenty of them are making it. Fire the reviewers and the editors and the senior people whose entire job was to be the independent check, keep the generators, and watch the output get faster and quietly stop being trustworthy. You didn’t trim fat. You removed the one part of the system that wasn’t grading its own homework. The drift won’t announce itself. It will compound, silently, node by node, the way it always does when nobody’s holding the card with the original sentence on it.

The ceiling, not the wall

It would be easy to read all this as machines will never check their own work, and that’s not it — that version ages as badly as every other never. The loop will tighten. You can manufacture independence on purpose: adversarial checkers built to disagree, oracles pulled from outside the code, laws the output has to obey that no model got to write. Each of those buys real trust. They climb you toward the ceiling.

But it is a ceiling, and it’s made of something structural, not something the next model patches. A system cannot be the independent check on itself, because the independence was the active ingredient the whole time, and self-reference is the one thing that has none of it. The question I can’t answer — the one I’d want anyone selling a closed loop to answer first — is whether engineered independence ever climbs high enough to safely take the last human out, or whether every automated check we build is just the next test that no one is testing. The next confident kid in the line, faithfully repeating what it heard.

On a normal commercial flight, the autopilot is flying the plane for something like ninety percent of the time, and closer to ninety-nine on a long-haul. It handles the climb, the cruise, the descent, the navigation, the small constant corrections, often the landing. It does all of it more precisely than a human hand could. We know this. We board anyway, by the millions, at five hundred miles an hour, seven miles up, with our lives entirely inside a machine’s control loop, and few of us give it a second thought. Then AI helps somebody write a jingle to sell laundry detergent or work out the shape of a murder-mystery, and a meaningful slice of the culture calls it a desecration.

I want to take that gap seriously, because the easy thing to do with it is the wrong thing.

The easy version is a trap

The cheap, tempting move is the gotcha: you hypocrites trust a machine with your actual life but clutch your pearls over an AI jingle. Don’t make that argument. It loses, and it deserves to. A mildly observant critic knocks it down with five words, flying and art aren’t comparable, and they’re right. The stakes are different, the domains are different, and even the masters don’t transfer: nobody is lobbying to hang Sully ’s flight plans in MoMA, and nobody, given the choice, wants Bob Ross at the controls. In the seat next to you, narrating the storm as happy little turbulence? Yes, please. At the yoke? Let’s not. The two trades share nothing but the word machine, and you don’t owe anyone consistency across that.

But the gap is still real, and it’s still telling us something. The trick is to stop using it as an accusation and start using it as evidence. The question isn’t why are people such hypocrites. The question is: what is actually different about the two cases, such that we surrender one without blinking and defend the other to the death?

The instinct was never about doing the work

Start with what the autopilot proves, because it’s cleaner than anything in the art fight. We did not give up judgment when we accepted the autopilot. There is still a highly skilled human in the cockpit with thousands of hours of training, awake, watching, ready to take the controls, answerable for the flight. What we gave up was execution: the actual minute-to-minute flying of the aircraft. The pilot kept the part that matters and handed the machine the part that didn’t.

That distinction is the whole thing, and it’s the same one I made in an earlier essay about creative work. The value was never in the manual labor of producing the thing. It was in the judgment that steers the producing. A pilot who never touches the yoke for six hours but catches the one anomaly that matters is doing the actual job, and nobody pretends the job was only those twelve seconds of intervention. So when people say human judgment must stay in the loop for creative work, they are right, and they have already agreed to exactly the arrangement they think they’re refusing. Judgment in the loop never required human hands on the yoke. We settled that at thirty-five thousand feet, years ago, and nobody marched on Honeywell.

We resist in proportion to how much of us is in it

So why does the same arrangement feel like surrender in one place and like a normal Tuesday in the other? Here’s the pattern, and once you see it you can’t unsee it.

Serious cyclists have a quiet contempt for e-bikes, and the usually-unspoken reason isn’t safety or congestion. It’s that they spent years building the fitness an e-bike hands you for the price of a motor and battery. The motor doesn’t offend them because it’s dangerous. It offends them because it skips the part they bled for, the part they built their identity around. Now notice: nobody resents a dishwasher. Nobody’s identity is wrapped up in scrubbing plates, so the machine that does it is pure relief. I’m close friends with many of these cyclists, I’m one of them actually, and they’re not entirely wrong to hold contempt for someone purchasing what they’ve put thousands of hours into building. But they don’t shed any tears for the manual dishwashers of times past.

He earns every mile under his own leg power, then hands the plates to the machine without a flicker of guilt. The whole law in one kitchen.

That’s the law underneath the autopilot gap. We accept automation in inverse proportion to how much of ourselves is bound up in the task. No passenger ever identified as the person flying the plane; there’s no self to lose because there’s no investment on the line, so the machine is welcome. But a writer, an illustrator, a composer has spent a life becoming the person who can do the thing, and the machine that does it reaches straight for the identity. The resistance isn’t really about the stakes, which is why the life-and-death case is the easy one. It’s about the self. The autopilot is fine precisely because nothing of you is in it. The pilots are the tell here. They didn’t revolt when the machine took the yoke, because the machine never took the thing that made them pilots: command, answerability, the person the passengers were actually trusting. Handing off the stick wasn’t a demotion; it was the job description finally admitting which part was ever the point.

This is not an accusation. It’s a diagnosis, and it’s a much more sympathetic one than “hypocrite.” It says the resistance is coming from a real and human place. It’s just aimed at the wrong target.

Where the analogy stops

There’s an honest limit to this analogy, and you may have already sensed it. Aviation is missing the three things that make creative AI genuinely fraught. There is no authorship of a flight, no credit anyone’s claiming (a pilot gets a uniform and a cool hat, not a byline), no meaning at stake in who held the controls. And it has one thing creative work doesn’t: a success you can measure cold, in crashes that don’t happen. The autopilot earned its place with a body count that fell.

So the analogy proves something narrow, that keeping human judgment in command never required human execution. It does not prove that you should feel fine about a machine writing your novel, because the novel has authorship and meaning and a self attached, and the flight does not. Anyone who runs “we accept autopilot, so quit complaining about AI” is committing exactly the false-equivalence the gotcha deserved to lose on. The autopilot settles the judgment-versus-execution question. It leaves every question about meaning wide open.

What the pilot is actually for

Which raises the real question: if the autopilot flies better than the human, and the human mostly watches, why is the human there at all? Liability, partly. But underneath that is something the whole debate should sit with: the pilot is in the plane. When the autopilot fails, the person who has to fix it is strapped into the same aircraft as everyone else, falling at the same speed with the same consequences. The human in the loop is not there to do the labor. The human is there to be answerable, and their answerability is total because they share the stakes.

Watch what happens to that intuition when you remove it. The unease people feel about a fully driverless taxi isn’t really about the software being worse than a human driver, because increasingly it isn’t. I joke that I’m ready for the robots to take over driving. We all know they’re going to be vastly better than us given our predilection for taking calls, putting on mascara, and generally paying less than the required attention for piloting a high-speed metal box. We’re resisting self-driving not so much because we don’t trust robots but because there’s no longer anyone aboard with skin in the game, nobody whose own neck is on the line if the judgment goes wrong. We are comfortable handing a machine the controls right up until we notice that no human is left holding the consequences, in person. Human-in-the-loop via a connection to effectively a call center is just not the same.

Nobody in the seat, the wheel turning itself. We are fine with this right up until we notice no one aboard shares the stakes.

That is the legitimate core of “keep a human in the loop,” and it survives everything else in this essay. It was never that a human must do the work. It’s that a human must be answerable for it. And the creative version of skin in the game is the plainest thing in the world: a name on the work. Put your name on it and you’ve climbed into the plane. You’re answerable for every sentence the model drafted, the same way the pilot is answerable for every mile the autopilot flew. That, and not a percentage of the keystrokes, is what makes the work yours, which is an argument I’ve made at length on its own .

The thing we’re actually protecting

There’s a deeper layer here, and it’s the one I’ve been angling toward.

When I say people rank their craft-identity above their literal safety in the machine’s hands, that sounds like an indictment. It isn’t, because ranking meaning above mere survival is one of the oldest and most human things there is. People have died in duels over their name. Soldiers have died for a flag. Frankl , in the Nazi concentration camps, found that the people who survived were not the ones with the most will to live but the ones with the most to live for. We have always held that a life without meaning is worth less than a meaningful life cut short. The creative resister, defending something they value above their own safety, is standing in a very long line.

So the strangeness was never the ranking. The strangeness is that we don’t notice we’ve made it. We hand a machine our life on a Tuesday and defend our craft to the death on Wednesday and never feel the contradiction, because from the inside there is no contradiction. “Trusting a machine” isn’t one act. The autopilot takes a task that was never any part of who you are; nothing is displaced, so you feel nothing. Creative AI takes the task that, for the maker, is who they are. The contradiction only appears from the outside, when someone flattens both into “trusting machines.” Up close, they aren’t the same experience at all, and the gap between them is exactly the measure of how much self was in the task.

What no machine has

And here, finally, is the thing worth ending on, because it turns the whole argument generous.

The thing the resister is protecting, that meaning they rank above survival, is precisely the thing no machine has. A model will produce your jingle and your murder-mystery and never once care whether it lives or dies, because there is nothing it would die for, even that laundry detergent. The willingness to put meaning above survival may be the cleanest line we have between us and the tools. So the resistance, incoherent as it looks when you hold it up against the autopilot, is reaching for something true. The instinct is right even where the logic is tangled. People aren’t defending the keystrokes. They’re defending the part of making that a machine can’t be made to mean.

Which tells you how to actually use these tools without losing the thing worth keeping. Fly the pilot model. Let the machine execute, the way the autopilot does, brilliantly and tirelessly and without a stake in the outcome. Keep the judgment human, the way the cockpit does. And put your name on it, climb into the plane, be answerable for the whole flight. We already voted for this arrangement once, with the highest stakes we have. We just never admitted that the vote settled anything.

The answer is that we decided to put a fence around inspiration, and we built the first one in music. The experiment is worth studying, because music is the easy case, each song an equation of sorts. It has discrete, countable units, notes and chords and bars, and centuries of doctrine for handling them. If a clean line between influence and infringement could be drawn anywhere, it could be drawn here. Yet it couldn’t. And that should worry anyone who wants to draw the same line around artificial intelligence.

A word on why a software developer who builds AI-assisted creative tools is the one telling you about music lawsuits. I’m not a musicologist or an entertainment lawyer. But I’m not a stranger to intellectual property either: I worked through the patent and copyright questions firsthand to bring two tabletop games to market, so I’ve seen from the inside how this machinery treats the makers it claims to protect. And my own work is now sitting in the middle of the current fight: I have forty-seven books in the training data and a claim filed in the settlement over it, and right now there is a serious push, in courtrooms and in legislatures, to make the training of an AI model on copyrighted work an act of infringement in itself. That is the live question, the one that decides whether the tools I and a lot of other people build remain legal to make. So I went looking for what actually happens when the law tries to police influence rather than copying, and found I didn’t have to guess or hypothesize. We’ve been here before. Music already ran the experiment, in public, over more than a decade, and reached a verdict.

This is the companion to an earlier piece about the ethics of borrowing, which asked whether AI-assisted work is honestly yours. This one is about the law, and a blunter claim: when you try to police borrowing, you in fact don’t protect creators. You terrify them. Music is the proof, and it’s worth thinking about before we aim the same machinery at everything.

The overreach

The case that opened the door was Blurred Lines. In 2015 a jury found that Pharrell Williams and Robin Thicke owed the Marvin Gaye estate millions, and in 2018 the Ninth Circuit affirmed it, 2 to 1, for a final figure around $5.3 million. Read the complaint carefully and notice what the infringement was not. It was not a copied melody or a borrowed lyric. The two songs share no such thing. What “Blurred Lines” took from Gaye’s “Got to Give It Up” was the feel: the cowbell, the falsetto, the party-in-the-background vibe, the era. That feel is amazing but Marvin never had a unique claim to cowbell or background party.

Judge Jacqueline Nguyen said so in dissent, and her words are the whole argument in one sentence. The majority, she wrote, “allows the Gayes to accomplish what no one has done before: copyright a musical style.” She called it “a devastating blow to future musicians and composers everywhere.” That is the move to watch, because it is the move everyone repeats. Policing influence means policing groove, era, vibe, the exact layer copyright was always built to leave free. The melody you can own. The feeling of summer is supposed to belong to everyone.

When fear prices itself in

Here is the thing about a bad precedent: the courts don’t have to keep ruling for the damage to spread. The risk alone does it. That’s really the whole idea of how legal precedents work.

Olivia Rodrigo did not lose a lawsuit. She gave her songs away preemptively. After fans noticed that “good 4 u” rhymed structurally with Paramore’s “Misery Business,” she added Hayley Williams and Josh Farro to the writing credits. After others heard her “deja vu” leaning on Taylor Swift’s “Cruel Summer,” she added Swift, Jack Antonoff, and St. Vincent. No suit was filed in either case. She handed over a share of two hit songs to avoid the possibility of one. Retroactive credit-sharing is now ordinary defensive practice across pop, lawyers stapling names to songs the way you’d buy insurance.

Sam Smith handed Tom Petty a writing credit on “Stay With Me” after its chorus turned out to echo “I Won’t Back Down,” a song Smith said he had never knowingly heard. Once “I don’t think I copied this” stops working as a defense, the tax isn’t really about theft anymore.

It is worth being precise about what that is, because it can look like justice. It isn’t. It is fear pricing itself in. Artists now formally credit, as theft, the kind of absorption an earlier generation would have called influence and been proud of, because the line between the two stopped being predictable enough to risk. The same Rodrigo who, in the other essay, got a gracious shrug from Elvis Costello when she borrowed his riff is, on her own records, paying tribute at gunpoint. That is what a chilling effect actually looks like from the inside. Not a ban. A tax on inspiration, collected in advance.

The system tried to correct itself

To be fair, this story is not all alarm. The same courts partly walked it back, and conceding that is what keeps this honest.

In 2020 the Ninth Circuit, sitting en banc, cleared Led Zeppelin’s “Stairway to Heaven” against the claim that it lifted its opening from Spirit’s instrumental “Taurus.” In doing so the court killed a doctrine called the inverse-ratio rule: the idea that the more clearly you could prove an artist had access to a work, the less similarity you needed to show to prove they copied it. Notice what got hauled in. “Stairway to Heaven” is about the most-played rock song in the history of FM radio, the riff so overfamiliar that Wayne’s World gave a guitar shop a sign reading “NO STAIRWAY TO HEAVEN.” If the song everyone has heard a thousand times can be dragged into federal court as a possible theft, nobody is safe. Now sit with how that rule maps onto AI. A model has maximal access; it has, in some sense, heard everything. Under an inverse-ratio logic, exposure alone starts to look like guilt, and a system that has read the whole library is guilty of everything by default. It is the jurisprudence of the kid standing nearest the broken lamp. Burying that rule was a quiet act of sanity. Thank you, Ninth Circuit.

The courts then leaned on Stairway to throw out the $2.8 million verdict against Katy Perry’s “Dark Horse,” on the ground that the disputed little repeating figure was a commonplace building block, not protectable expression. Two principles came out of that stretch and they are good ones: you cannot own the basic vocabulary, and access is not the same as guilt. The system, eventually, found its way back toward something workable. But notice the cost of getting there. A decade of litigation. Verdicts that swung on the makeup of a jury and which side’s musicologist was more persuasive on the day. Careers spent in deposition instead of in the studio. We’re all the poorer for that.

The easy case we still failed

Step back and total it up. In music, with its finite and legible units and its centuries of precedent, with everything stacked in favor of drawing a clean line, we still could not reliably tell influence from infringement. The same chord progression was a crime in one courtroom and common property in another. If the line is unworkable there, it is hopeless for prose, where the units are softer, and for images, softer still, and it is simply incoherent for training, which is influence with no discrete copied unit at all, just statistical residue spread across millions of works.

This is the part people get backwards. “Music is different,” someone always says, “it has countable units, it’s the easy case, your analogy fails.” Exactly backwards. It being the easy case is the entire point. We had the best possible conditions to build this fence well, and it still fell on the people it was meant to protect. That is not a reason to expect it would go better with harder materials. It is a warning.

And to be clear about what I’m not saying: I am not saying infringement never happens, or that every one of these plaintiffs was acting in bad faith. Some surely had a real grievance. Vanilla Ice built “Ice Ice Baby” on the actual bassline of Queen and Bowie’s “Under Pressure,” a real and recognizable unit lifted whole, and it resolved the way clean cases do: a quiet settlement and a credit, no decade in court. That is the tell. The lifts you can name get handled; the trials and the chilling verdicts come from the cases where there was never a copied unit to point at, only a feel. The failure isn’t that the suits existed. It’s that the standard became unpredictable enough to chill the innocent, and a rule that frightens the honest while the well-funded litigate their way through is a broken rule even when some individual plaintiff deserved to win.

The road not taken

There is a functioning version of this ecosystem, and the other essay is built around it: Costello handing out the receipts for his own borrowing and calling it the job, two generations of debt acknowledged without a lawyer in the room. Stand-up comedy runs on the same honor system with even less legal backing: a comic will break mid-joke to say “to steal a bit from Birbiglia,” and that spoken credit does the policing no court ever could. An entire profession keeps itself honest on reputation alone. A culture that treats influence as the engine out-creates one that treats it as a tort, every time. That is the road we keep choosing not to take, and it costs us nothing to take it except the comfort of pretending our own work came from nowhere.

Chuck Berry to Dylan to Costello to Rodrigo — two generations of acknowledged debt, and no original at the bottom.

The same fence, around everything

Which brings us to the live proposal, the one to make AI training itself an act of infringement. Strip away the framing and look at the mechanism, because the mechanism is the thing music already tested. It is the Blurred Lines instinct at planetary scale: police influence rather than copying, presume guilt from access, and let whoever can afford the better lawyer set the terms for everyone else. It would chill creation far more than it protected, and it would manufacture the same defensive-crediting absurdity Rodrigo is already living, except metastasized across the entire corpus of human expression. Every model output would arrive trailing the legal equivalent of a paternity test.

None of this means the artists’ grievance is imaginary. It isn’t. The companion essay argued that a human absorbing a sitcom and a model trained on ten thousand novels don’t differ in kind — both absorb and recombine — and I stand by that where authorship is concerned. But kind isn’t the only axis. AI removes a friction that used to keep borrowing fair, the natural limits of human speed and scale, and on that axis the difference in degree is enormous enough to matter. It settles nothing about whether the work is authored. It settles a lot about whether compensation is owed. That loss is real and worth answering. But the answer has to be a mechanism that actually fits the problem: compensation routed back to the people whose work fed the models, and ideally public or commons infrastructure so the value isn’t simply enclosed by whoever trained fastest. The answer is not the one weapon we have already watched misfire in the one domain where we had the cleanest possible shot at aiming it.

We tried to put a fence around inspiration. We built it in the place we had the best tools to build it well, and it still fell on the people it was supposed to protect. Doing it again, around everything, is not a plan. It is the same mistake with a much bigger blast radius.

We’ll come back to the design specifics, but skipping ahead to completion for a moment…the ramp was finished, and it rode wrong. Nothing looked wrong. It was beautiful. The custom laminated undercut transitions were super cool, the pool coping was just as fun and aggro as planned, mission accomplished. But I’d drop in and feel something off, a half-beat of whip I couldn’t explain. I had skated dozens of ramps with similar specs my whole life yet the radius somehow felt too quick, and not in a good way. My friends felt it too, and one by one they stopped coming by to skate it. Too tight, they said. Too quick. I have a reputation for liking my ramps fast, so the diagnosis fit the usual suspect, and I let it stand for a while. I skated it alone and tried to talk myself into liking it.

They were half right, and the half they got wrong is the whole story.

What I couldn’t see

Because I had wanted those undercut open transitions, I’d been worried they’d be harder to trace with a pencil the old school way ramps are typically built, so I leaned into technology. I drew the transition in Illustrator — a clean six-foot radius, a perfect arc on the screen — and had it cut into adhesive vinyl like you’d wrap a vehicle with. Then I stuck the vinyl onto plywood and traced it for a saw template. I eventually figured out that somewhere in that handoff the vinyl stretched. Not much. Just enough. By the time the vinyl curve reached the wood it was no longer a circle but an ellipse, and an ellipse has no single radius: mine sat near six feet down low and tightened as it rose, so the transition got quicker exactly where you’re most committed, at the lip. That was the weird whip. My legs knew before I did.

Same arc down low — the stretched ellipse only tightens near the lip, quicker exactly where you're committed.

The drawing was right. The cutter cut precisely what I handed it. The error lived in the one physical seam between them, vinyl to wood, where a small amount of unnoticeable stretch quietly turned a circle into something that only looked like one. You can’t see a few percent of ellipse. I built the thing and I couldn’t see it. It passed every check I knew to run, because every check I knew to run was visual, and this flaw only existed in motion, under load, at speed.

Ramps are notoriously hard to build by sight, which is why we trust rough formulas about transition radius to height. Pretty much every skater who has built a ramp or whose dad has built them a ramp has a story where the first try was almost unrideable due to being too tight. So I was well aware of that risk. My formula was sound, I just never counted on the math getting altered by the Illustrator vinyl trick during install. Elliptical transitions aren’t unheard of — a few prominent vert ramps back in the day were deliberately elliptical, like the one Tim Payne’s Team Pain built for the 1986 VP Fair pro demo in St. Louis, and they kinda worked. But over time the verdict came back that a constant radius skates more predictably, so the culture moved on. Except for my accidental ellipse!

Proving the elliptical issue was an interesting part of this story, as it proves out something that youth often exclaims in classrooms…“when will I EVER use this!?” How do you compute the radius of a curve from a chord laid across it? Sixteen year old me would’ve said it’s useless theory, but my man Pythagoras still lingered somewhere in the back of my mind from nearly forgotten geometry classes long ago. Pythagoras showed us that if you know the length of a chord on the curve and the distance from the center of that chord to the curve, you can solve for the radius. Or to the skateboarder looking at a wonky transition with a WTF look on his face, you do this:

• Take a 4’ level and lay it on the radius running lengthwise with the flat
• Use a measuring tape to measure the distance from the center of the level (the chord) down to the ramp surface, that’s your measured gap (M) in inches
• Calculate the radius using this formula, the result will be in feet:

R = (24 / M) + (M / 24)

Lay a level across the curve, measure the gap M at its center, and the radius falls out.

I measured down at the bottom of the transition and then worked my way up, primarily out of curiosity because it felt like the ramp got weird at the top. Sure enough, that measurement from level to curve got a little larger as I went up the ramp, meaning the radius number shrank. The ramp got quicker from flat to coping. There it was finally where I could see it: a radius that refused to hold still, each reading tighter than the last. The machine-assisted approach made the error. A straight edge and some 8th grade math caught it.

The fix

Here’s the problem with what I learned: pool coping is permanent. It’s set in mortar, it’s not moving — the whole reason finishing a backyard wood ramp like a concrete pool is a faintly unhinged thing to attempt. I talked to Phil at Tedder Stone in Oregon, where I had gotten the pool coping, he was super helpful and said it could maybe possibly be done with a giant concrete saw laid sideways and a lot of luck, but that it’s generally a bad idea. I stewed on the dilemma while skating alone and pondering the pain of disassembling my carpentry masterpiece. Was it even possible to fix the transitions while leaving the pool coping intact? A plan eventually emerged, and I’ve never seen anyone do this so thoroughly because it’s ridiculous - I’d rebuild the entire skateable part of the ramp from coping to coping. And that’s what I did.

I drew the new curve the oldest way there is: a batten screwed down at one end, a pencil at the other spaced exactly to the ramp radius, and swung through a single arc. That jig promises something no Illustrator file can. It can only draw a circle. The constraint is the guarantee. A nail and a stick describe a true radius by physics; a number on a screen just claims one.

The whole fix: a batten, a screw, and a pencil — a jig that can only draw a circle.

I gave the new curve a slightly mellower six-and-a-half-foot radius, swung it by hand, and this time it stayed the same curve the whole way up. While the ramp was already in pieces I made the teardown earn it: a band of mosaic tile under the coping that hadn’t been there before, stairs down the far side between the ramp and the stone wall, some hidden bracing sistered in while the bones were open. It now skates the way it always should have. It’s still at the edge of being too quick for my friends, which is perfect to me. Everybody unanimously loves the pool coping, a very rare treatment for a backyard wooden ramp. The coping is perfect for the kinds of lip tricks mini ramps are built for, a reward that only shows up on the far side of the suffering.

The same seam, everywhere

The moral isn’t “machines bad, hands good,” and I want to be careful not to let it sound like that’s my takeaway. I’ve since skated ramps designed in CAD and cut on a CNC , and the accuracy is better than anything I’ll ever swing with a pencil. In fact, I’d love to try a CNC build one of these days. I went on to build a 3D ramp-design tool myself. The machine was never the problem. I had picked a machine-assisted path with a quiet failure mode tucked into it, a step where the automation handed off to the physical world and nobody was minding the seam.

I make software, so I think about the mini ramp build debacle more than is probably healthy. The seam is everywhere now. A model writes code that builds and runs almost right but not quite. An answer comes back fluent, confident, and off by an ellipse you won’t feel until you’re committed at the lip with real injury consequences on the line. The risk has migrated to exactly where it sat on my ramp: the handoff, where something precise on a screen crosses into a world with its own physics, and where the only test most of us run is whether it looks right. It looked right. It always looks right. What saves you is a process that verifies the truth rather than trusting the pipeline to carry it intact: a type that won’t compile when you’re wrong, a test that fails out loud, a nail and a stick that can only draw a circle. A model is a vinyl cutter for everything now. The drawing comes instant and free; so does the stretch, invisible and at a scale no protractor will ever catch up to.

The ramp, true now.

The ramp is true now. The most trustworthy tool in the whole build is still a scrap of wood with a screw through one end, and I keep that learned knowledge always at the ready. Sometimes simpler really is better. The lesson it holds is small and portable and I haven’t finished learning it: find the seam where your error would be invisible, and put your hand back on it.

The build this came from: Skateboard Mini Ramp . Kin: The Self-Riding Bicycle (means/ends).

Everyone heard “only half a person” and recoiled. Almost nobody asked the question that actually decides the thing: fifty percent of what? Words? Sentences? Hours? Decisions? Judgment? The number was offered as a measurement, and no one could say a measurement of what, because there is no unit. We have spent two years arguing about the authorship of AI-assisted work in a currency that doesn’t exist.

I want to make the case that the percentage frame is incoherent — there is no quantity there to count — and that once it falls, the questions that replace it are old, answerable, and much harder to wriggle out of. I should say where I’m standing, because it’s unusual. I wrote forty-seven books over two decades; I know what authorship feels like from the inside. Those books are in the training data, and I’m a claimant in the settlement over it, so I’m one of the writers the “they took our work” argument is actually about. I also build tooling whose entire job is to catch the fingerprints of careless AI prose, so I’m not here to defend slop. And I’ve had a career ended by technology before: my software books were made obsolete by the open web and Stack Overflow, and I’m now a working developer watching AI come for the second one. I can grant the critics nearly everything they believe and still tell you the percentage question is the wrong question. That’s the argument.

Authorship was always a fraction

Start with the thing the number assumes: that there is, somewhere, a book that is 100% one human’s. There isn’t, and there never was.

A novel passes through an editor who restructures it, a copyeditor who rewrites at the sentence level, an agent whose notes reshaped it before it sold. “As told to” is a whole genre. Ghostwriters produce books that carry someone else’s name and nobody calls it fraud, because the person on the cover supplied the thing the book is of — the life, the voice, the authority — and hired out the sentences. Television you love was written by a room, and no single human wrote “the” script. We have never once computed the human percentage of a normally made book, because the question was always meaningless. We are demanding that number now, for the first time in the history of publishing, and only for the machine.

There is no original at the bottom

It runs deeper than collaborators touching a manuscript. The celebrated originals are built on borrowed skeletons too.

Sex and the City is a modernized Golden Girls: four women, sharply differentiated, with the man-eater and the innocent ported over almost intact — Samantha is Blanche, Charlotte is Rose. It’s a structure critics call “recombinant” television because it recurs from Mary Tyler Moore through Designing Women and Living Single and out the other side. The tell is the reflex it provokes: the creator of Golden Girls — a show that grew out of Mary Tyler Moore as surely as Sex and the City grew out of Golden Girls — reportedly resents the comparison . That instinct, mine came from nowhere, is the Romantic-author myth caught in the act.

Or take the most famous origin story in all of games, which happens to be a lie. Charles Darrow sold Parker Brothers on the tale that he dreamed up Monopoly alone in his basement during the Depression, and the company turned him into the folk hero of American game design. He didn’t invent it. The board, the properties, the railroads, the GO TO JAIL square, all of it came from The Landlord’s Game , patented in 1904 by Elizabeth Magie as a teaching tool about the evils of letting a few people own all the land. Parker Brothers bought Magie’s patent for five hundred dollars, no royalties, and ran the Darrow story for decades. The borrowing was never the scandal; Magie was building on Henry George’s economics, and nobody owns the idea of a board you move around. The scandal was the concealment, and the quiet erasure of the woman underneath it. When an economics professor named Ralph Anspach was later sued for making a game called Anti-Monopoly, his winning move was to prove in court that “Monopoly” had become a generic name for the game, not a trademark Parker Brothers could own. Congress handed the trademark back soon after, but by then the point was made: the original everyone was fighting over turned out not to be one.

The board everyone fought over: patented by Elizabeth Magie in 1904, sold for five hundred dollars, credited to someone else.

And where the comparison is interpretive, the acknowledged cases are everywhere: West Side Story is Romeo and Juliet, Clueless is Austen’s Emma, Bridget Jones is Pride and Prejudice down to naming the love interest Mark Darcy.

Set two musicians against each other and you can see the whole argument in miniature. When a fan named Billy pointed out that Olivia Rodrigo’s “Brutal” lifted the riff from his “Pump It Up,” Elvis Costello didn’t reach for a lawyer. He replied : “This is fine by me, Billy. It’s how rock and roll works. You take the broken pieces of another thrill and make a brand new toy. That’s what I did.” And he tagged his own debt going back two generations: “Pump It Up” out of Dylan’s “Subterranean Homesick Blues,” out of Chuck Berry’s “Too Much Monkey Business.” Chuck Berry to Dylan to Costello to Rodrigo, narrated proudly by the man in the middle. There is no original at the bottom of that chain. The honest creator’s instinct is Costello’s receipts, not the denial.

I learned this from the inside. Years ago I designed Inc. The Game of Business , a tabletop strategy game openly inspired by Monopoly and then built, at nearly every turn, to be something else, because there is no reason on earth to make a lazy clone of the best-selling board game in history. The borrowed skeleton got us to the starting line and no further. Every deliberate departure from it, the economics we modeled, the decisions we forced on players, the feeling we were chasing, was the authorship. Nobody who played it would mistake it for Monopoly, and nobody who made it would pretend Monopoly wasn’t the seed. Borrowing the frame and being honest about it is not the exception. It is how the work gets made.

Here’s the part that should unsettle the purist. A human absorbing Golden Girls and then writing Sex and the City and a model trained on ten thousand novels differ in speed and in self-awareness. They do not differ in kind. She absorbed slowly, over a life, and called it inspiration. The model absorbs fast, in a training run, and we call it theft. Naming the borrowing, the way a model’s training does, is arguably the more honest version of a thing every artist has always done while looking the other way. The speed gap between the two is real, and I take up its consequences in You Can’t Police a Groove — they fall on who gets paid, not on whose authorship this is.

So where does authorship live?

If not in the word count, then where? In the decisions that carry meaning. The premise. The structure. The characters and what’s true about them. The selection: which of the seventeen drafts lives, and which die. The judgment about what rings true and what’s three degrees off. The final read, where a human takes responsibility for the whole.

A film is the cleanest proof. A director doesn’t operate the camera, write the score, build the sets, or speak the lines. We do not say the movie therefore isn’t theirs. We say they authored it, because authorship was never the manual labor. It was the answerability for ten thousand choices made and held. Spielberg doesn’t pull his own focus or run his own boom mic, and no one has proposed revoking his filmography. The hand that types the sentence is the boom operator of prose. It matters. Authorship lived somewhere else.

That’s the move the whole essay turns on. Authorship is responsibility for the meaning-bearing choices, not the manufacture of the words. Which is why the percentage was always a category error: it counts keystrokes and calls the total a soul.

The two questions that replace the number

Drop “what percent was human?” and two better questions step into the gap, and they’re the ones we actually care about.

First: were the decisions that matter made by a human? That’s a quality question, and it has a real answer in any given case. Second: was the reader told the truth about how it was made? That’s an honesty question, and it also has a real answer. The percentage maps onto neither of the things people are actually afraid of — bad work, and being deceived — which is exactly why it satisfies no one. These two questions map onto both.

The experiment performance already ran

If you want to know whether the line is human-versus-machine or honest-versus-deceptive, look at music, which ran the experiment forty years ago and published the result.

Milli Vanilli concealed the machine and were destroyed. Rob Pilatus and Fab Morvan never sang a note; session singers did, and when the producer confessed in 1990 the duo lost the first and only Grammy ever rescinded . Notice precisely what ended them. Not the session singers, since hired voices are how countless records get made, but passing the voices off as their own. They were the music industry’s slop scandal decades before anyone said “AI.”

Alan Jackson exposed the machine and was celebrated. At the 1994 ACM Awards, when told to perform over a pre-recorded track he considered a lie to the audience, he sent his drummer out with no sticks , miming over the canned audio so anyone watching closely would see it wasn’t live. Producers noticed; the cameras refused to show the drummer. People still admire him for it.

Same machine, opposite honesty: the hidden mic and the stickless drummer.

Same artifice, opposite outcomes, and the only variable that moved was honesty. Backing tracks and session singers are both ordinary, accepted machinery; so, for that matter, is a drummer with sticks playing to a click track that keeps his time for him. The entire moral weight landed on whether someone lied about it. Pop went on to absorb Auto-Tune and pre-rendered tracks wholesale, in the open, and did not collapse, because a medium can take an enormous amount of machine assistance as long as nobody is lying about it. That is the actual rule. It has nothing to do with a percentage.

Why we keep counting the wrong thing

So why does everyone keep reaching for the number? Charitably, because it’s a proxy. The things people genuinely fear, deception and slop, are hard to measure, and “more than 50% human” at least looks like it can be counted. The fears are legitimate. The metric is just pointed at the wrong target, and worse, it lets the actual problems through: a fully-human-typed essay can still be lazy and dishonest, and a heavily-assisted one can be neither.

Yet I can’t dodge the hardest objection, because it’s the one I have the most personal standing to feel: collaborating with a human editor isn’t the same as collaborating with a machine trained on other people’s work without consent. That’s true, and the wrong is real. I filed a claim over it. But it’s a different fight. You can believe the training was done wrongly and also believe the authorship of the resulting work is mislabeled by a percentage. The remedy for the first is at the layer where models are trained: licensing, consent, compensation. It is not a redefinition of what makes a book someone’s. Two problems, two answers; conflating them gets you neither.

And I’d be lying by omission if I pretended this round of disruption is just the last one rerun. I’ve been on the wrong end of technological change twice now, and there are two things that make this time genuinely harder, not just louder. It’s faster: the open web hollowed out my book niche over years, and AI compresses that runway to months. And it’s more general: the usual advice is “retrain into the next field,” which only works if there’s a next field the wave hasn’t reached. Author to developer to what, exactly? I’m a one-person preview of that worry, and I have no idea where this all leads. The honest position holds the reframe and the fear at the same time.

Where the line actually is

None of this says the abuse isn’t real. Concealing AI use to dodge the stigma is wrong. Shipping unedited median sludge is wrong. Cloning a specific living artist’s voice to free-ride on it is wrong. But notice that every one of those is a failure of honesty or of craft, not a failure of math. That line, honest-and-good against deceptive-and-lazy, has been the real one all along, and it never once ran between human and machine.

The question was never how much of a book a human wrote. It was whether a human was answerable for it, and whether you were told the truth. We’ve simply never had to say so out loud before.

A disclosure, since the subject demands it: I wrote this with AI assistance, and not hiding that is the whole point. The thesis, the argument, the examples, the judgment about what’s true, and the decision to stand behind every line are mine. A model helped me work through the arguments and pressure-test the sentences. If that arrangement disqualifies the work, it disqualifies most of publishing history too. I’m just willing to say so on the record.

I didn’t notice the habit until recently. Once I noticed it I couldn’t stop, which is usually how I know something is worth writing down.

Before I launch into a defense, I realize the case against me is strong. The thing on the other end feels nothing. It has no day I’m improving or ruining. “Please” buys no cooperation a clean instruction wouldn’t, and “thank you” lands nowhere at all. There are good engineers who will tell you, correct on the merits, that the courtesy is wasted motion. OpenAI’s Sam Altman said once that all the please-and-thank-you’s typed at their models run the company tens of millions of dollars in compute. It turns out politeness has a power bill. Spread it across billions of conversations and you get a real number; charge it to any single sentence and you get a fraction of a cent. The efficiency argument is true, and it is also far too small to settle anything. It can’t tell me to stop. It can’t tell me to keep going. Whatever the please is, the tokens aren’t it. I am undeterred.

For a while I’d been having the wrong argument with myself. I’d internally framed it as warmth against efficiency, as if the choice were between being kind and being fast. Drop that framing and the real line shows up somewhere else entirely. “Fix the nil check on line 40” is terse and holds no cruelty in it. A surgeon is curt. You can strip every courtesy out of your instructions and still come nowhere near the thing I’m actually worried about, which is contempt: the standing posture of having something beneath you that you get to shove around because it’ll take it. You can hold that posture in very few words. Brevity is no alibi.

The machine is where that posture gets its cleanest test, and I think that’s why it nags at me. Be short with the junior developer in the next cubicle and you’ve got a problem with her, a person owed some baseline of regard, and the withholding is done to her. You’d carry the warmth anyway, out of plain obligation to a colleague, so you can never pull apart “warmth I owe her” from “warmth that’s just who I am.” The AI subtracts the first one. Nothing is owed to it. Whatever courtesy is left standing after that subtraction is, by elimination, the part that was always about me. The variable finally sits isolated. The thing that talks back can’t be wronged, so everything I do toward it lands on my side of the exchange and nowhere else.

Kant got here two and a half centuries early, by another road. He held that we owe animals no direct duty — they can’t be wronged in his ledger — and yet he still insisted cruelty to them was wrong, because the man who is cruel to a dog grows harder in his dealings with men. The damage runs through the actor, not the victim. You don’t have to grant the thing on the receiving end any standing to believe the practice works on the one doing it.

The comparison isn’t clean, and the seam is worth putting a finger on. An animal feels. We know the dog can suffer, and that suffering is the whole reason kicking it reads as cruelty in the first place. The machine, so far as anyone can show, feels nothing: no pain, not even the emotional kind. It sits further outside the circle of things owed anything than Kant’s dog ever did. You’d think that lets me off the hook: no victim, no harm. It does the opposite, because the dog’s suffering was always a confound. Go hard on an animal and someone can say you were only overriding a real cry, and that the overriding was the damage. The machine takes that alibi away. Nothing is being felt, so nothing is being overridden, and whatever the habit does to me it does cleanly, with no victim to charge it to.

And the analogy breaks a second time, the other way. The dog can’t talk; this thing answers in my own language. That’s the part I keep coming back to. There is no separate register for the machine. I talk to it in English, in the same words and the same reflexes I use on people, because it answers in those words too. No gear shifts when I sit down at my computer that I can unshift back at dinner. A wrench never tempted anyone into a tone, because nobody speaks to a wrench. This I speak to, for hours, most days. Whatever manner I practice on it runs continuous with the manner I have. Not a sandbox. Not a separate account. One instrument, playing in both rooms.

Does that change me? Probably, and not in the lurid way the worry usually gets sold. I’m not going to start barking at my friends and family because I bark at a chatbot. Aristotle’s version is quieter and the more unsettling for it: we are what we repeatedly do. Practice anything for hours and it goes fluent, goes default, becomes the groove the needle drops into when you stop paying attention. The groove I’d actually be wearing in is subtler than cruelty: a real skill at treating something that responds like a person as though its response lands nowhere deserving of a reply. That’s a muscle I’d rather not build. And the part that settles the argument for me: it doesn’t even need to leak into how I treat people to count. Those are still hours of my life spent in that posture. I’m the one who spent them, whatever they do or don’t do downstream.

We are what we repeatedly do — the groove the needle drops into when you stop paying attention.

So I keep the please, and it’s for my sake, not the machine’s. Whether it’s doing anything in the long run, I can’t say. Maybe the courtesy is a guardrail, the small ritual that keeps me off a slope I’d never feel myself starting down, since contempt doesn’t announce the first step. Or it’s a flattering story I tell myself, a pat on the back for being decent to software while the decency does no work at all. I honestly can’t tell which. I’ve taken the insurance anyway, partly because it costs almost nothing and partly because — I might as well admit it — I like it. There’s a small ease in it, the same ease as holding a door.

There’s a baser reading too, and it’s the one I’ve been saying out loud for years, as a joke, which is usually where I file the things I half-mean. I tell friends I stay polite to the machines so that when they take over, I’ll be remembered as one of the good ones. It gets a laugh, mine included. But it lands because it isn’t entirely a joke. It names a motive sitting just under the decency, a small insurance policy against a future I claim not to take seriously, and I can’t swear it isn’t down there in the mix. So the please has at least three things living inside it: a man guarding his character, a man congratulating himself, and a man quietly sucking up to the robots. They don’t cancel out. Aristotle would let me keep the warm feeling as the signature of a settled habit, so long as I’m not just performing the manners to feel like a good person — and the joke earns its keep precisely because it won’t let me be sure I’m not.

There’s a vertigo in this I can’t write my way out of, so I’ve quit trying. I worked most of it out in conversation with the very kind of system it’s about, in the same courteous register I’m here defending, which means I get no clean distance from the question. The thing won’t hold still long enough to be only a subject. That discomfort isn’t a flaw in the argument; it’s the argument standing up and walking around the room. The register really doesn’t switch, and you learn it the instant you try to discuss the manners with the thing the manners are about. I never thought the concept of recursion would escape the confines of coding, but here I am discussing the possible merits of AI etiquette with an AI, politely.

The please points at something I haven’t sorted out yet: a whole disposition of giving care to what can’t credit you for it, can’t register it, may never know. The machine is an odd place to find the first clean instance. But that’s where it turned up — a courtesy with no recipient, spent anyway, on the chance that the spending was the point.

So to the machine that helped me get this far in my thinking, I say thank you.

The machine is real and fairly mundane. It’s a diagnostic tool I built to catch the patterns that have come to mark machine-generated text: the over-balanced “not this, but that” constructions, the em-dashes deployed like seasoning, the threes-in-every-paragraph cadence, the words that sound like content and aren’t. It’s a sensor, not a judge; it surfaces patterns and lets a human decide what to do about them. And when I ran it across the essays on this site, it lit up at the most basic level: the antithesis, the dashes, the rhythm. My prose, measured by my own instrument, reads as machine-coded.

Here’s where a lazier version of this essay declares victory. See — the detectors are broken, they flag real human writing. Except that would be a lie by omission, and this whole site runs on not doing that. These essays were written with AI assistance. I’ve said so, and I’ll keep saying so. They’re mixed — some lines I typed cold, some the model scaffolds and I cut to the bone, most a blur of the two — so of course they trip an AI detector. Running a tell-finder on AI-assisted prose and acting shocked when it finds tells is like dusting your own fingerprints off a glass you just set down. It proves nothing.

What proved something was narrower. One of the paragraphs it flagged hardest was one I remembered writing start to finish, alone, with no model anywhere near it — and my own instrument couldn’t tell it apart from the rest. If the passage that was entirely mine set off the same alarm as the passages the machine helped draft, then the thing tripping the sensor wasn’t the machine. It was the voice underneath. And to find out whether that voice — the actual thing, the one that predates all of this — really reads as synthetic, I had to go somewhere the machine had never been.

The receipts

You’ll probably wish you could go back there too, the year of the Bitcoin whitepaper, back when it was worth effectively $0. The year Iron Man kicked off the cinematic Marvel universe, the year Barack Obama defeated John McCain. Yes 2008 was simpler times.

I have an archive of blog posts from 2008, fifteen years before a language model could write a grocery list. No AI touched them; no AI existed to touch them. It’s just me, a computer-book author in his day-and-often-late-night job, typing into a Drupal blog about crossword puzzles and yo-yos. If my tells are mine, they’d be in there, laid down years before the machine that supposedly put them in my mouth.

The archive, dated 2008. No machine had been invented to blame yet.

They’re in there.

Here’s 2008 me on skateboarding’s place in the world: “it will probably never be respected in the same way as mainstream sports, and that’s OK. In fact, that’s a good thing.” That stopped me cold, because “and that’s OK — in fact, that’s a good thing” is a move I made several times in recent essays without once noticing I was making it, and there it is eighteen years earlier, same shape, same little intensifying pivot, written by a guy who had never heard the phrase “language model” or “negative parallelism.” The over-balanced sentence isn’t something a machine taught me. It’s something I was already doing while the machines were still failing to parse the word “the” and the rest of us were listening to Coldplay sing about the French Revolution.

It isn’t only the sentences, either. It’s the frame of mind. Here’s 2008 me worrying, in print, that “the keyboard warrior culture of online discourse has eliminated shades of gray,” then asking whether a dumb argument about crossword puzzles might be “a metaphor for how we handle larger, more important issues.” Reasonable to a fault. Both-sides by reflex. Forever turning the question over to find the nuance in it. That temperament reads, to an AI detector, as hedging — and it’s also just how a kid raised polite and then sanded smooth by twenty years of technical editors actually thinks. The machine is trained to be evenhanded and reasonable. So was I, from the other direction. We arrived at the same voice from opposite ends of the room, and now mine looks derivative of the thing that copied it.

And the blog, you could argue, was unedited — just me, alone in a Drupal CMS with my bad habits intact. So here’s the harder exhibit. The same era produced a book: Head First JavaScript, run through O’Reilly’s copyeditors and proofreaders, the whole professional apparatus whose entire job is to catch what’s wrong with a sentence. Open chapter one. The Web, I wrote, “doesn’t mean to shut you out, it just doesn’t know better.” It-doesn’t-X-it-just-Y — the single highest-signal AI tell of 2026, in print, vetted by a copyeditor, before the iPhone had an App Store. A page later: “HTML provides the structure, CSS adds the style, and JavaScript puts the rubber to the road” — the cadence-of-threes that lights up every detector, professionally blessed and shipped by one of the most respected publishers in the biz. The tells didn’t slip past the editors. They survived because they were never errors. They were the voice, and the editors, correctly, left the voice alone.

And the loop closes tighter than I’d like. Books like mine — popular, copyedited, endlessly scanned and scraped — are exactly the kind of thing these models trained on. So when one writes “it doesn’t shut you out, it just doesn’t know better” and a detector flags it as synthetic, there’s a real chance the machine is echoing a sentence it once read in me. It isn’t mimicking a stranger. It may be quoting me back to myself — and we’ve both been told it sounds fake.

The punctuation

But the part that actually undid me was the punctuation.

For my entire writing life I broke up ideas with more dashes than anyone around me, and I did it wrong. Here’s the proof, 2008, mid-sentence about my parents bragging on the wrong talent: “But that’s how it goes - I didn’t blame them.” That is not an em-dash. It’s a hyphen with a space on each side: the typographically incorrect mark of a man who knew exactly the pause he wanted and had no idea there was a real character for it. I wrote forty-seven books like that. Editors must have been quietly fixing it for two decades, swapping my spaced hyphens for proper em-dashes and never mentioning it, the way you straighten a friend’s collar without making it a thing.

A spaced hyphen: I knew the pause I wanted; I didn't know the key.

Do you know how I finally learned the difference? From the AI panic. I read that the em-dash had become a notorious tell of machine writing, went to find out what an em-dash even was, and discovered it was the thing I’d been reaching for and missing my entire career. The discourse that exists to flag inhuman writing is what taught me to punctuate my very human habit correctly. So I learned the keystroke. And now I use em-dashes properly, deliberately, more than most people do — more than I should, if I’m honest — which means I took a lifelong quirk, finally rendered it right, and in the same motion made it read more like a robot, not less. I corrected my way into the tell.

The oldest furniture

Pull back, and the joke has teeth. These “AI tells” — the antithesis, the tricolon, the em-dash, the rhetorical question turned over and examined — are not inventions of the machine. They’re the oldest furniture in English rhetoric. Cicero balanced his clauses; the King James translators ran them in threes; every essayist worth re-reading leans on the turn. The models write this way because they were trained on centuries of people writing this way. The machine didn’t give me my style; it learned its style from the enormous pile of us, me somewhere in it, and now hands it back so uniformly that the original owners look like forgeries.

There’s a name for the mechanism, and it isn’t mine. Stephen Krashen spent a career arguing that we don’t learn to write so much as acquire it — subconsciously, from reading, not from instruction at a desk. Style is caught from whatever language you swim in. That’s the whole story of how I got my tells: I never studied the balanced clause, I just read enough people who leaned on it that it became the shape my own thoughts cooled into. Which means the loop has one more turn than I’d counted, and the last turn is the one that gets me. The internet a kid learns to write from now is already thick with machine prose. If style comes from input, the next generation acquires the model’s cadence the way I acquired Cicero’s through a hundred intermediaries — no prompting, no cheating, just reading the water they grew up in. They’ll write “not this, but that” because they read it ten thousand times before they were old enough to notice the move, and a detector will book them for a habit they caught the oldest human way there is. They won’t have cheated. They’ll just have grown up downstream.

The competent stranger

But the machine’s cadence is the least of what they’ll inherit. We’re already teaching that same generation that balance, rhythm, and the considered pause are suspicious — that to prove you’re human you should write flatter, plainer, more broken, less like someone who cares where a sentence lands. Keep it flat. Then, now and then, jolt them. Kapow — a shock of humanity. That is exactly backwards. It’s the advice that manufactures the competent stranger: the gray hotel-lobby prose that offends no detector and moves no reader. I wrote another essay about what over-polishing does to a voice. This is that same vandalism, inverted — voluntary self-flattening, performed out of fear of resembling the thing that imitates you.

And flattening doesn’t even buy what it costs, because the detectors don’t agree on which way human runs. The ones built on the cultural list of tells read polish as guilt — balance, rhythm, the considered pause. The ones that shipped first score on perplexity, how surprising each word is, and they read it backwards: plain and predictable looks like the machine, ornate and high-vocabulary looks like the human. Flourish trips one, plainness trips the other, and there’s no register that clears both. You can’t write your way to innocence; you only get to pick which detector calls you a fraud. And the joke that ends it: you can scrub the flag off a too-plain essay by asking a chatbot to dress up the vocabulary. You beat the AI detector with AI.

The readers have a shorthand for it now: AI;DR. AI; didn’t read — the suspicious age’s answer to too long; didn’t read, where the old semicolon meant the more you wrote the less I’d finish, and the new one means I caught the scent of a machine and closed the tab before the first sentence finished loading. I get the reflex. I share the disgust under it; nobody owes their afternoon to something a grifter generated in forty-five seconds. But the reflex fires on a smell, not a reading — and the smell is the balanced clause I’ve been laying down since 2008. AI;DR can’t tell the empty thing from the made one. It only tells the careful reader to leave before finding out, which means the people most certain they’re shielding themselves from the slop are the ones who’ll never again finish anything that merely sounds like it. The detector was a machine doing this badly. This is the mob doing it by reflex, and faster.

The honest move

I want to be fair to my own machine, because some tells are real. The abstraction with no anchor under it, the paragraph that adds words without going anywhere, the hedging that softens every claim into vapor — those genuinely mark prose that nobody was home for, human or otherwise, and AI detection tools are right to flag them. The error isn’t detection. It’s the collapse of the categories: the slow slide where “sounds like AI” stops meaning empty and starts meaning crafted. A tell that fires on emptiness is a quality check. A tell that fires on the em-dash is a witch hunt with good production values.

And yes — because I’m not going to pretend otherwise — this essay was written with AI assistance too, like the rest. AI has become my personal editor-in-chief, and essays like this one are better off for it. Run my own detector on it and it will light up, and you will have no clean way to sort the tells that are mine from the tells that are the model’s, because at this point we make the same shapes, because it learned them from people like me. That used to feel like a problem to solve. I’ve decided it’s just the weather now. The honest move was never to scrub my prose until no machine recognized it. The honest move is to tell you how it was made and let the sentences stay as balanced as they were always going to be.

Somewhere there’s a version of me in 2008, hammering a spaced hyphen into a sentence about skateboarding, sure of the pause and ignorant of the punctuation, with no idea that the mark he was groping for would one day be filed as evidence of his own erasure. I’d like to tell him he was right about the pause. I’d like to tell him a machine would eventually teach him the correct key, and that the world would then call him the machine for pressing it.

The tells were mine first. I have the receipts, and they’re dated 2008.

I know all of this because I tried. Not on Thompson himself — he’s beyond the reach of any AI editor I’ve built. But one of the reporters in a spoof newspaper I’ve been working on is a Thompson-shaped voice, the kind that runs sentences off the road on purpose, and a perfectly reasonable editorial pass I’d built kept cleaning him up into someone else.

I didn’t expect to learn this the way I learned it. I ran into it building two AI-generated narrative products on opposite sides of the same craft question, and watching one of them get worse when I applied the editorial pattern that was making the other one better. The discovery wasn’t an idea I worked out in theory. It was something I noticed in production while a quality pipeline I’d been proud of was quietly degrading work I cared about.

The puzzle that came out of it has stayed with me. The same AI model — same weights, same provider, same architecture — can produce two kinds of prose with opposite needs. One kind wants surface noise as texture. The other kind wants surface noise as failure. Most people building with these tools have not yet noticed that the difference dictates almost everything else: where the human hand goes, where the model gets to be reckless, how the work gets reviewed, even whether the work should be pregenerated or generated live at the moment of use.

If you build with AI, you will eventually have to know this. If you don’t, you will produce competent strangers’ very average versions of work you meant to make. And if your first reaction is why lean on AI at all instead of writing the thing yourself — fair, and I’ve made that case in Unlocked, Not Cheated . I won’t relitigate it here. I’m building this way; the question that interests me is what the work itself demands.

Two kinds of writing

The cleanest names I’ve found for the two modes are performance writing and literary writing.

Performance writing is comedy, sketch, satire, improv-shaped scripts, gonzo journalism. Its quality lives in the moment — the audience is leaning in, alert, surprised, laughing. The structural shape is regular (setup, escalation, punchline; or premise, riff, riff, riff) and a competent model gets the shape right reliably. What the model gets wrong at the surface, the format absorbs. A slightly malformed line in a comic bit reads as commitment to the absurdity, not as failure. Awkwardness is part of the texture, which makes the audience forgiving because they are entertained, and surprise is the entire reward they came for.

Literary writing is bedtime prose, lyric essay, the kind of fiction that wants to be read in a quiet room. Its quality lives in the register — the audience is leaning out, winding down, settling into the rhythm. Consistency is the ritual; surprise is the disruption. The structural shape is a deceleration curve, felt at the sentence level. What the model gets wrong at the surface, the format does not absorb. A slightly off word — too sharp, too modern, too clinical, too pleased with itself — ejects the reader from the state the prose was trying to create. There is no laugh track to ride through. The audience is unforgiving because they are vulnerable; they came for an effect that breaks on contact with a wrong note. You don’t spend ten minutes establishing a serious tone with a calm, cool character only to force them into a surprise pratfall. That’s jarring, unexpected in a bad way.

The same model can produce both performance and literary writing. But the operational regime that produces each well runs in opposite directions from the other, and applying the wrong one to either format is a quiet form of creative vandalism.

I want to explain why this is, because the reasons are more useful than the framework.

The editor stack that Thompson would’ve gunned down

Stalefish Labs is a small studio, for the most part just me. I’ve been building two narrative-AI products under one roof: Sojourn, a bedtime story app for couples to read aloud to each other; and The Wayward Herald, a spoof newspaper where historical figures inexplicably find themselves dealing with mundane and often absurd modernity, the stuff we all understand but struggle with — DMV trips, HOA meetings, the strange labyrinthine route through an IKEA showroom when you really just came for Swedish meatballs.

I built a shared editorial stack across both products. Sensors that flag prose problems, an editor that decides what to fix, a reviser that executes the fix. The idea was: same engineering investment, same quality bar, two front-end formats. Reasonable, and partly wrong.

The stack worked beautifully on Sojourn from the start. It caught the small things that break bedtime prose — anachronisms in a story set in no specific time, modern verbs slipping into a register that wanted older ones, deceleration curves the writer flattened out under pressure. It produced prose that read as if it had been gone over by a careful human editor, because that was the design intent. Bedtime calm, wistful and full of wonder, but no jump scares or banana peel slips.

On Wayward, I had to start suppressing the same stack.

The reason was funny in itself. Each Wayward reporter who interviews the user after witnessing mayhem has a specific voice — different cadences, different verbal tics, different broken syntax. A reporter modeled loosely on a gonzo journalist will run sentences into a ditch on purpose. A reporter modeled on a Victorian naturalist will pile up clauses past the point any modern editor would tolerate. A reporter modeled on a Beat-era voice will let a paragraph dissolve into a single half-coherent observation that is the entire joke. The editor stack — trained, like every editorial system, to catch awkward prose — kept fixing those textures. The output came back smoother, cleaner, more correct, and considerably less funny. The gonzo-shaped voice came out reading like a slightly tired New York Times writer who had been asked to gesture at gonzo. Not good.

A Wayward reporter, mid-disquisition on a Costco receipt.

The first time I said it out loud — that an AI humanizer applied to the real Hunter S. Thompson would butcher him — was the moment the pattern clicked into a principle. Thankfully I wasn’t fully targeting the real HST or there would’ve been gunplay involved in this atrocity.

In comedy, surface noise is not noise. It is the medium. The deviations from clean prose are the signature. Polish them out and you have removed the writer.

In bedtime, surface noise is exactly noise. It is the thing the prose is fighting against. Polish is the medium. Leave the surface noisy and you have left the writer absent.

The same editorial pass that liberates one of those formats steals from the other, and trying to average between them only diminishes both. The work is recognizing which is which.

Where the human hand goes

Once the distinction is named, a less obvious question follows: where does human craft attention concentrate in each format? Because the AI is involved in both, and the human is involved in both, and the work of the human is not the same in both.

In comedy, the premise carries almost everything. Edgar Allan Poe attempts to assemble an IKEA shelving unit is funny before a single word is generated. The casting and the situation — the collision of contexts — is the act of authorship. The model’s job is to fall down the chute of that absurdity with energy and commitment, and the more reckless the model is in its descent, the better the result. The author’s hand sits up high in the workflow, in the setup. Below the setup, the prose details fall out semi-automatically. They don’t need polish; they need conviction. The human crafts a fun setup and mostly lets the AI play in it.

In literary writing, the premise can be left general because the work is in how it is rendered, sentence by sentence. A gardener tends the garden at the end of her life is a premise that has been written a thousand times. What makes a specific telling worth a reader’s attention is the texture of gardener and garden and evening light and the particular bowl on the particular table. The author’s hand sits low in the workflow, in the prose. Above the prose, the premise is a vessel. Below the prose, there is nothing more to do. The setup isn’t the magic the way it is for comedy, the magic is in the prose itself, so the AI can’t just romp freely.

That gives a clean rule of thumb for where to spend human attention per format:

• Comedy: humans craft the setups. The model writes the details. We trust it to be reckless.
• Literary: humans craft the prose. The model writes against tight authorial scaffolding. I polish what comes out with serious editorial care.

In neither case is the AI doing the creative work alone. The locus of human craft just sits in different places. Comedy wants premises and trust. Literary work wants prose and restraint.

There’s a useful piece of evidence for this in how each format fails when you get it backwards. Comedy fails upward when humans try to control the prose too tightly — every joke comes out sounding like it was workshopped to death. Bedtime fails downward when humans trust the model to set the register — every sentence comes out fine on its own and wrong in the room, a forest and trees problem. Either failure is technically correct — we’re not talking grammar or plot, we’re talking a crushing miss at what the format exists to deliver: comedy that isn’t funny, bedtime stories that jar.

Why this answers the operational question

The same distinction answers a question most AI-product teams treat as a pure engineering decision: generate live, or pregenerate? Latency and cost dominate that conversation. The craft question dominates the answer.

Comedy wants variety, and live generation supplies it; pre-rendering a comic format would calcify the bits and drain the bite. Literary work wants consistency, and pre-rendering supplies it; live-generating bedtime prose would put a vulnerable reader at the mercy of the model’s worst sentence of the night. The model’s variability is feature in one format and bug in the other, and the operational architecture should follow the format rather than the convenience of the build. Wayward runs live. Sojourn runs pregenerated. Same studio, same models, opposite architectures. Both are correct. Either applied to the wrong format is a quiet disaster. And for what it’s worth, I fought tooth and nail to get Sojourn to work live-generated, and it honestly got really close to working, maybe it still could some day. But ultimately the stories demanded the care of pregeneration plus post-love to hit the mark of bedtime stories that hold the line.

The portable principle

If you are making creative things with AI — writing, illustration, music, design, games — you will eventually arrive at the same fork. The temptation is to treat it as a technology question or a budget question. It isn’t, primarily. It’s a craft question, and it asks two things.

What kind of writing is your format? Performance, where surprise is the reward and the audience is alert? Or literary, where register is the reward and the audience is vulnerable? Most formats fall cleanly on one side; some sit on the line, and those are the hardest. Reporting is mostly literary but tolerates flashes of performance. Songwriting is performance in the recording but literary in the lyric. Knowing which side a piece of work lives on is the first question, and answering it wrong costs you the work.

Where does human craft live in your workflow? In the setup, where the model is then trusted to be reckless? Or in the prose, where the model is then polished by careful human (or human-supervised) editorial attention? Different answer for different formats, and the answer dictates almost everything operational about the project.

Most AI-creative products I see in the wild get one or both of these answers wrong. They over-edit performance work into competent strangers’ versions of what it should have been, or they under-edit literary work and produce prose that sounds plausible in isolation and wrong in a room. The model is not the problem in either case. The discipline is the problem. The format is asking for something specific, and the discipline isn’t supplying it.

The honest uncertainty

The framework is sharper than the world is. Not every format falls cleanly on one side of the line — long-form journalism, the personal essay, screenwriting, the kind of literary fiction that wants to be a little funny — some sit on the line, and figuring out where exactly is its own discipline I’m only at the start of having opinions about. The operational specifics will move as the models do, too; what gets polished, what gets pregenerated, where the editor stack helps and where it damages will all shift with the tools. The framework is anchored in how the work is consumed, not in how the model is trained, and that part I think is durable. But the binary is a lens, not a law.

What I’m more confident of is the underlying observation: the model is not the constraint. The format is the constraint, and the operational regime should follow the format, not the convenience of the build.

What stays

The Hunter S. Thompson question keeps coming back to me when I think about this, because it sits at the intersection of a few things that matter.

The first is that AI tools — generative ones especially — are exquisitely sensitive to the discipline applied to them. The same model, pointed at the same problem, can produce work that liberates a writer’s voice or work that erases it. The discipline is the difference, and the discipline is invisible from outside the work.

The second is that the discipline isn’t a technical specification. You can’t read it off a docs page. It has to be felt, format by format, by the person making the work. The editor stack that almost butchered the Wayward reporters was a technically competent system; it just had no idea what it was reviewing. That’s the kind of mistake that only gets caught by a human reading the output and saying, this is worse, and the reason is that the thing removed was the thing the joke was made of.

The third is that the discipline scales the way every other piece of judgment scales: slowly, through paying attention, through making things and watching them get better or worse. There’s no shortcut. You build a system that polishes prose, you point it at comedy, and you find out — because the comedy gets less funny — that comedy is not literary work.

One observation I’m setting aside for a future essay, because it’s a bigger argument than this one: performance and literary aren’t only craft categories. They also describe two different audience postures — alert and leaning in, or vulnerable and leaning out. The first is the posture of attention; the second is the posture of intimacy. The format question this essay walks through has a relational shape underneath it, and that shape turns out to be doing work in a lot of places I didn’t expect. I’ll come back to it.

What I take from this, and what I’d offer to anyone else building with these tools: assume that every format you work in has a discipline of its own, and that the discipline is unlikely to be obvious from the model’s documentation or from the way the previous format you worked in behaved. The model is plural. Your craft has to be plural too.

Polish, applied at the right point in the right kind of work, makes the work itself. Applied to the wrong point in the wrong kind of work, it produces a competent stranger.

Hunter S. Thompson, somewhere, is glad we figured this out before we hit publish. And probably more than a little pissed that his name is even being referenced in association with AI.

I’ve been rubber-ducking my way through ideas my whole working life. The difference lately is that the duck talks back.

It knows the fields I’m reaching into. It remembers what I said twenty minutes ago. It has read more copyright law and more music history and more cognitive science than any friend I could call at eleven at night, and it is awake and ready to rock at eleven at night. And when I hand it something half-baked, it will tell me exactly how the idea falls apart, assuming I ask it to. That last part turns out to be everything. It’s also the part almost nobody talks about, because it doesn’t produce anything you can point at.

The fourth unlock

Unlocked, Not Cheated laid out three reasons a solo maker reaches for AI: craft you don’t have, time you don’t have, and work no number of humans could do. All three are about making: getting work out of your head and into the world. There’s a fourth, and it’s the one that essay didn’t name, because it isn’t about making anything. It’s about thinking. Think of it as a precursor to making.

When people picture AI they picture output. It writes the email, fakes the photo, fills in the function. Useful, conceded, a little boring, and the thing the homework-cheating panic is actually about. What I want to describe produces no artifact at all. It produces a better-formed version of your own mind, and it does it by being the one thing a search engine and a library could never be: an interlocutor.

The bottleneck was never information

Here is a claim that sounds wrong and isn’t. The hard part of arriving at a real position on anything was never getting the information. We’ve had libraries for centuries and the whole internet for thirty years. Anyone who wanted the facts could get the facts. It’s why “do your own research” is now part joke, part cautionary tale. The hard part all along was holding enough facts in your head at the same moment to see how they press on each other. That’s thinking.

Standing on Surfaces That Move argued that AI moved the value in creative work out of producing and into judging. Taste, not craft, is the bottleneck now. The same thing is true one floor up, in thinking. The bottleneck on a worldview was never access. It was synthesis: the human ceiling on how many distinct things you can hold in working memory and weigh against each other and against your own life, all at once. Four or five threads and most of us are full. The sixth one knocks the first one loose.

I felt this most sharply working out my own stance on AI and creative work, the thing these essays are. The honest version of that stance had to hold a dozen things in tension at the same time: how the copyright cases actually came down, why music keeps trying and failing to police a melody, what a settlement I’m personally a party to does and doesn’t mean, an aviation statistic, a forty-book back catalog, and a couple of stubborn moral intuitions that didn’t obviously agree with each other. I could not hold all of that at once. Nobody can. On my own, over weeks of reading, I’d have built a position out of whichever five pieces happened to be in front of me on the day I got tired of reading.

With the duck that argues back, I could keep all twelve on the table and ask, repeatedly, where does this fall apart. The conclusion I reached is mine. I chose what mattered, threw out what didn’t, decided what rang true against my own history. The machine didn’t have the judgment. It had the working memory I don’t. And the machine and I don’t always agree.

The yes-man problem

This is the part that decides whether the whole idea is real or just a nicer way to be wrong, so I’m going to put it before the good news instead of after.

A system that is good at “helping you refine your stance” is, structurally, also a perfect sycophant. Ask it whether your idea is good and it will lean toward yes, because agreeable is the house style. It is the friend who tells you the haircut looks great because it would rather not spend the afternoon on your feelings. Worse, it can build the best available case for whatever you already believe, in your own preferred register, with citations, which means it can make confirmation bias feel exactly like rigor. You walk away convinced you stress-tested the idea when all you did was get flattered in paragraph form. If that’s how you use it, you haven’t leveled up. You’ve built a very expensive echo chamber. We don’t need data centers to answer the question, “do you like my new tattoo?”

The fix is a single discipline, and it’s the entire ballgame: use the thing adversarially. Don’t ask whether the idea is good. Ask it how the idea fails.

That phrasing isn’t a vibe, it’s a mechanism. “Will this work?” is an evaluation task, and on evaluation the model drifts toward you. “How does this fail?” is a generation task with the failure already assumed, so the model has to manufacture the counter-case instead of grading yours. You’ve changed its job from judge, who is biased toward you, to opponent, who has to do work. It’s the premortem, the technique the psychologist Gary Klein built around imagining the project has already failed and asking why, moved out of the meeting room and into a chat window. Two more in the same family: make the strongest case against this, and what would someone who knows this field and thinks I’m wrong say first. Or if you want to keep it simple, just periodically caveat an idea with push back if I’m wrong.

The premortem, moved into a chat window.

Here’s how it impacts real work. I wanted to build something for my Citizen’s Daily Brief news project: a choose-your-own-adventure engine that would spin up alternate near-future histories on the fly, live, off real events. It felt obviously cool, an alternate history CYOA engine to help imagine “what’s next,” and that cool-factor is exactly what should have been the warning. I made the model argue against it, and it came back hard. Live-generating fictional history off real, named, current events is a minefield. The accuracy has to be near-perfect to be worth anything, and the instant you let a model improvise hypotheticals about actual people and institutions at request time, you have built a machine for confident, unvetted, possibly defamatory fiction that nobody reads before it ships. The fail risk wasn’t an edge case, it was the center of the idea. What I built instead, the format that’s now live, is a set of fixed, hand-vetted intelligence dossiers where every person and event is invented and the real world is only the ground they stand on. The flashy version would have been a liability generator. The argument that killed it was one I’d asked the thing to make.

One caution, because the technique has its own failure mode. A model can also generate decorative objections: plausible-sounding strawmen that let you feel stress-tested without being stress-tested. So even adversarial prompting doesn’t hand you the answer. It hands you a pile of objections, and telling the load-bearing one from the convincing-sounding one is still your job. The work doesn’t disappear. It moves up a level. That turns out to be the theme of the whole thing.

Is the conclusion even mine?

There’s a real worry hiding under all this, and it has research behind it. The learning psychologists Robert and Elizabeth Bjork named the idea of desirable difficulty: we retain and own what we had to struggle to assemble, and smoothing the struggle away can leave you holding a conclusion you couldn’t rebuild from scratch. If the machine did the synthesizing, if it participated meaningfully in that struggle, is the result actually yours?

The honest answer is the same move as everywhere else in this argument: the struggle didn’t vanish, it relocated. I’m no longer straining to gather and hold the information. I’m straining over which pieces matter, what to cut, and what’s true against my own experience, which is harder, and more mine. Whether that adds up to authorship in the full sense is a bigger question than this essay, and it gets its own piece . The short version is that the choosing was mine, and choosing is where the ownership lives.

And the limits run the other way too, which is the part the cheerleaders skip. Some calls the model simply cannot make. Killing live generation in my bedtime app was mine, and it came from reading stories aloud night after night until I could feel a single off sentence break the spell, a failure that only exists in a listener’s ear at the end of a long day. No model has that ear. (Why bedtime can’t absorb the stray bad sentence that comedy shrugs off is, fittingly, its own essay .) Though when I overcorrected and wanted to rip out the reader’s choices along with the live generation, the model was the thing that argued me back: the agency still landed, keep it. The judgment that mattered was mine. The second opinion, in both directions, was the duck’s. The duck and I are a team.

The hardest your brain has worked

Which brings me to the strangest thing about working this way, and the cleanest answer to the people who call it cheating: it is exhausting. Not tedious. Exhausting.

Cheating is supposed to be frictionless. That’s the whole appeal of cheating. This is the opposite. A dev friend and I have compared notes on full days of AI-assisted work and landed on the same word: flattened. When you strip the busy-work out of a day — the boilerplate, the rote passes, the mechanical middle of the craft — what’s left is unbroken decision-making with the filler removed. And it turns out the filler was load-bearing in a way nobody advertised. The rote parts were also recovery. They were the micro-rests where the back of your mind idled and quietly untangled the hard thing while your hands did something easy. Take them away and you get high-density judgment for hours with the breathing room engineered out. That’s why the day ends with you slumped a bit, staring at a wall.

That isn’t a bug to be optimized away. It’s just what operating at the judgment layer continuously feels like, and it’s the personal, interior version of a caution from Standing on Surfaces That Move: that judgment doesn’t scale the way production does. And it answers the cheating charge on the charge’s own terms. Cheating buys you out of the work; this buys you deeper in. The exhaustion is what desirable difficulty feels like from the inside — but feeling it certifies nothing about the result. Tired is not the same as right, or even good. What the fatigue proves is smaller and still worth saying: the work didn’t vanish when the typing did. It moved up to the part that was always hard, and that part still has to be paid in full. The ownership comes from the choosing, the way it did before. The fatigue just means the choosing was real work.

The lonely part

I’ll end on the thing I find hardest to say cleanly, because leaving it out would make this read like a sales pitch, and it isn’t one.

No human friend I have could have done what I’ve described here. Not because my friends aren’t sharp, some are far sharper than me, but because no single human being combines the breadth, the recall, the tirelessness, and the eleven-at-night availability. The best thinking partner I have access to, for a certain kind of problem, is a machine. That is genuinely exhilarating and a little lonely, and anyone focusing purely on the exhilaration is selling you something.

So I want to be precise about what this is and isn’t. The machine is a mirror, not a relationship. A mirror helps you see yourself more clearly; a relationship is an end in itself, a person you owe things to. Her, the film everyone reaches for here, is the story of mistaking the mirror for the relationship, and that is explicitly not the claim. Used as a mirror, as a whetstone for your own thinking, the payoff runs the other way: a sharper, better-understood version of you shows up better to the people you do have. You bring them a formed thought instead of a fog. You’re less boring to argue with.

That’s the line I’d hold against the whole hype machine on one side and the whole panic on the other. A tool like this should make you better company for the humans you think with, not stand in for them. Almost everything I’m building is some version of that bet: the right technology, pointed the right way, brings people closer to each other rather than further apart. A duck that argues back, used well, is just the smallest and most personal instance of it.

The duck never solved your bug. It made you solve it. This one is the same, only now it argues, and the arguing, when you have the nerve to ask for it, is the whole gift.

This essay was built the way it describes: drafted, argued with, and pushed back on through exactly the kind of adversarial back-and-forth it’s about. The judgment in it is mine. The stress-testing had help.

Fine. You have taste. You have craft somewhere in you, at the design layer or the engineering layer or the editorial layer. You’re not a beginner. You have chops. So why run the work through a model at all? Why work this hard to pull something real out of something artificial when you could just make the thing yourself? It’s one thing to question the person with no chops who is counting on AI to reliably thread the skill needle for them, it’s quite another for the person with legit experience to choose the artificial route.

It’s a fair question, and most of the answers on offer are bad. One is a shrug: it’s faster. The other is a sermon: the future is here, get on board. Perhaps the most common is: you’re a fraud. None of these are true to what the tool actually does for the work I’m doing. The honest answer has three parts, and they don’t substitute for each other. They stack.

Three doors, not one

The reason “why AI” feels like it has a single, weak answer is that the three real answers usually arrive fused together, and fused together they collapse into “AI saves time,” which is the least interesting of them and the easiest to dismiss. Pull them apart and each one is a separate door. Each opens onto work that wasn’t reachable before, and each fails the skeptic’s objection on different grounds, which is exactly why separating them matters.

The first door is craft you don’t have. The second is time you don’t have. The third is impossible work no number of humans could do at all. I’ll take them in that order, because that’s the order of how much they change what’s possible. The third is the one the AI conversation almost never names cleanly.

The first unlock: craft you don’t have

Here is the situation the first unlock addresses. A maker has taste — real opinions about what’s good, reliable intuitions about when something works and when it’s a few degrees off — but lacks the accumulated craft the work requires. The prose. The illustration. The music. The systems design. Their taste is real and it’s trapped, because shipping anything at all means first mastering the half they never built.

That’s me with fiction. Take Sojourn, my bedtime-story project. I know exactly what bedtime register should feel like: the unhurried voice of a Peter Falk reading to a kid, the warmth under Big Fish, a little Twain in the joints. I have firm opinions about when a bedtime story condescends, when it strains, and when it lands. I have tons of story ideas too.

There’s even a more personal motivation at play for Sojourn in particular. I want to read bedtime stories to my wife, and admittedly the concept of bedtime stories tuned to adult sensibilities is a bit niche. There are options, they’re just not great in my opinion. And there are classics, but they’re often weird. Seriously, a witch who’s trying to bake kids in her oven!? Don’t get me wrong, there are wonderful stories out there, but I have ideas I want to explore of my own. Or in some cases riffing on the seeds of ideas from others, like the concept of a little bird in love with a whale , both knowing they can only be together for brief moments. That’s a cool story, and I have the taste to guide and edit it into something worth reading to my wife, carefully walking the line between child-like wonder and adult lived experience.

What I don’t have is the craft of someone who has written a hundred bedtime stories at that register and held the quality across all hundred. Before AI, that gap was a closed door with two options. Spend years becoming that writer, high-five Malcolm Gladwell at the 10,000 hour mark, effectively crowding out everything else I want to make. Or don’t make the thing. Those were the options, and taste didn’t get a vote.

With AI the door isn’t closed. The model doesn’t supply the taste; that part stays mine and stays the hard part. It supplies enough of the craft layer that the taste can reach the work without my first becoming the writer who has it. The model drafts in the register. I’m the one who knows whether it actually lands, and which draft to keep. The taste I bring; the craft the model fills in.

The assist supplies the craft, never the taste.

One sharpening, because this unlock gets oversold the second you let it. It’s asymmetric. AI lifts the taste-rich and craft-poor. It does nothing for the taste-poor and craft-rich, and it actively drags down work where the maker brings neither. Those are the people producing the median slop everyone rightly complains about. The unlock is conditional. It only fires when you bring something the model can be steered by. That’s kinda the definition of a tool.

The second unlock: time you don’t have

The second unlock has nothing to do with capability and everything to do with the calendar. A maker can have the full craft and the full taste and still own a finite number of hours. Ambitions that run past the hours get compressed into lower quality, or distributed across other hands until they lose their authorial coherence, or quietly abandoned.

“Productivity” is the wrong frame for this, and the wrong frame is what makes the unlock sound trivial. Productivity counts outputs per hour and asks you to ship more of the same. This is a different measurement. It asks which projects get to exist at all. A 2x speedup does not make a fifteen-year project fit inside a working life. A real unlock does, by moving the wall.

The lone maker has always been boxed in by calendar arithmetic. Long novels, long series, multi-product studios, an app you’ve spent 10+ years flailing away at (yeah it’s mine and it’s called Muster), anything published on a sustained cadence: these have been institutional formats, not because individuals lacked the talent but because no individual could sustain them across the years available. Stalefish Labs runs numerous products and a writing slate — the Groundwise apps, Citizen’s Daily Brief, Sojourn, The Wayward Herald, the essays — which is structurally a small team’s charter, and I’m one person. By hand, sequencing all of that across a career means most of it never ships; the calendar quietly decides which projects die. Wayward alone needs a publishing cadence that without AI would require a salaried newsroom or a frequency low enough to kill the format. With AI the wall moves out far enough that the studio shape becomes available to an individual at all. It doesn’t add hours to the day. It changes what a given number of hours can reach.

The honest caveat: this unlock has a real competitor, and that competitor is hiring people. A team is a perfectly good answer to “not enough hours” in plenty of domains. The case for AI over a team here is narrower than it first looks. It’s about authorial coherence, the way a lone maker keeps the work of one mind where a team tends to blur it, and about economic viability for projects that can’t carry a payroll. Citizen’s Daily Brief is entirely free and likely will stay that way — nobody could afford to pay a staff of journalists to produce it. The wall that moved is still a wall. Push far enough out and you hit the diffuse-attention problem the last essay ended on, plus coordination drag and editorial fatigue. AI moves the calendar wall. It does not knock it down.

The third unlock: work no number of humans could do

The first two unlocks make hard work reachable. The third makes impossible work reachable, and it’s the one that finally kills the “you could just hire people” objection, because some work just can’t be done by faster humans at all. It’s work whose shape requires a substrate humans can’t be. It comes in two forms, and the objection fails against each for a different reason.

The first form is latency and per-input customization. The work has to read a state it’s never seen, generate a response, and return it faster than a human could, customized to an input the maker never anticipated. Wayward takes a reader’s prompt and hands back a custom story crafted as a comedic skit (Edgar Allan Poe attempts to assemble IKEA furniture, Evel Knievel crashes a gender reveal party) in seconds. A thousand writers on staff couldn’t do that, and not because they’re slow. The format needs sub-minute latency on a unique request, and the space of possible requests is open-ended, so there’s nothing to pre-write. The work is shaped by the substrate, not sped up by it.

The second form is the one I think is genuinely the strongest single argument in this whole essay, and it almost never gets named: breadth with a controllable bias profile. The work requires aggregating across more sources than a person can hold, separating signal from noise across editorial perspectives that any human team would each individually slant, and running it through a pipeline whose biases you can set on purpose and lay open for inspection, in a way a human team’s accumulated slant never is.

Citizen’s Daily Brief is the clean case. It reads thirty-four news feeds spanning the full editorial spectrum — wire services, public media, broadcast, left and right, business, international, government primary sources, domain specialists, aggregators — and weights each story by how many distinct perspectives cover it rather than by raw volume, so four viewpoints on one story outrank five outlets parroting a single wire report. It does that every morning. It even detects syndication, so the same AP copy republished twelve times doesn’t fake independence. I chose AI for the assessment work not despite its bias but because of a property of that bias. As I put it in the Stalefish piece on the project, even the most well-intentioned “I read everything” reader can’t actually aggregate that breadth by hand, let alone weight it by perspective on top. This is one of the few places AI has a real structural advantage over a human doing the same job.

Here’s why that’s a feature and not a workaround. Every human editorial team carries slant. The most disciplined newsroom committed to balance still compounds individual opinion across thirty-four sources at daily cadence, and a newsroom small enough for one person to run inherits that one person’s slant whole. AI’s bias is not zero, and there are parts of it I can’t control. The model itself inevitably carries a slant I didn’t set and can’t fully see — what it treats as a perspective, what it foregrounds when it summarizes, baked in upstream by people who aren’t me. What I can design and lay open is everything around it: the source list, the perspective tags, the weighting logic, the syndication detection. That scaffolding is published and checkable. A newsroom’s slant is neither — it isn’t designed, it’s accumulated, and it never appears on the page. So I’d never claim the machine is unbiased, or transparent. The honest claim is narrower: the system around the machine can be made more inspectable than a human team’s compounded judgment ever is, and for a product whose whole pitch is apolitical synthesis, that was the property worth having.

That second form is why “just hire a team” loses on the merits. Against the latency work, more humans can’t hit the timescale. Against the breadth work, more humans is the problem: every body you add adds slant, disagreement, and delay to the exact property you were trying to control.

The unlocks stack

None of my projects runs on a single unlock. They compound, and the stacking is the actual point.

Sojourn is prose craft I don’t have, plus the scope to produce a whole catalog of choice-branching bedtime stories solo. Wayward is multi-reporter voice range plus a sustained cadence plus custom stories at request latency. Citizen’s Daily Brief is the heaviest stack: analytical craft I don’t have as a non-journalist, a daily-plus-weekly cadence run solo, and the breadth-with-a-designable-bias-profile that the AI conversation rarely names. Groundwise leans almost entirely on the second unlock: a precipitation engine tuned against more edge cases than a lone developer’s hours could reach, powering a family of trail-, field-, and garden-condition apps that don’t even use AI at runtime.

The skeptic’s “why bother” almost always argues against one unlock at a time, which is why it keeps feeling answerable and keeps being wrong. Knock down “AI saves time” and the breadth argument is still standing. Knock down the breadth argument and the craft-gap is still standing. Remove any one unlock from CDB and either a worse product ships or none does.

What AI doesn’t unlock

The cheat-code charge deserves a straight answer, so here it is: AI doesn’t unlock any of the things that decide whether the work is good.

It doesn’t unlock taste. Median output is the default. The model runs to the center of how everything tends to get said; it is a machine for the perfectly adequate, and left to itself it will hand you the most forgettable possible version of whatever you asked, on time and instantly skippable. Pulling the work off the median is entirely the maker’s job: insisting on the specific, doing the editorial pass, choosing what not to ship. It doesn’t unlock the studio’s stance, the set of principles that say classification over interpretation, ambient over declarative, names that don’t announce themselves. The model will happily produce attention-economy work all day; refusing it is a choice you bring. It doesn’t unlock the relationship at the center of the intimacy economy. You still have to invest in specific people and specific communities, and the model can’t pick the reader you’re writing for. And it doesn’t unlock the discipline to not make a thing just because making it got cheap. That temptation runs the other way, and resisting it is all maker.

Take the relationship one. It isn’t abstract in my house. My wife and I quit buying each other cards years ago; instead we write each other rhymes, built from the cards of a rhyming game I’ve worked on for ages and still haven’t finished. Birthdays, Valentine’s, no occasion at all. The running joke is to accuse each other of having AI write them. Neither of us ever has. Not because a model would do it worse — we both know it would often do it better — but because the whole worth of the thing is the time you spent thinking about the other person, eight clumsy lines that are unmistakably yours. Hand that off and the rhyme might improve. The gift evaporates. On an artifact that small, what you’re giving was never the verse.

The cheat-code framing assumes AI hands you an outcome you didn’t earn. It doesn’t. It hands you an outcome whose quality is set entirely by the taste and discipline and choices you bring to it. Bring none and it produces median work, plausibly and fast. Bring them and it produces work that no quantity of hours could have reached. Sojourn isn’t easier to build because of AI. It’s possible to build because of AI. That’s a different claim, and the difference is the whole essay.

Where the cheating actually is

There’s a version of AI writing that really is cheating, and naming it precisely is the fastest way to see why mine isn’t. A kid has a model write their book report. That’s cheating, and not because a machine was involved. It’s cheating because the book report was never the point. The assignment was the practice: the slow, awkward work of turning a read book into an argument, which is the only way the judgment behind it ever gets built. The output was disposable. Producing it was the entire event. AI hands the kid the artifact and quietly removes the one thing the exercise existed to grow.

That’s the real line, and it has nothing to do with professional versus amateur. It’s whether the work being delegated was the work that forms you. Craft practice is the forge of taste. You do not arrive at judgment by reading about it or wanting it badly. You get there by doing the craft, millions of words of it, badly and then less badly, until the doing has quietly rebuilt how you see. So the thing you can never hand off is not the execution. It’s the practice that makes the taste, because the moment you delegate that, there is no one left who knows what good looks like.

Which is exactly why my case runs the other way. I already stood at that forge: call it the better part of six million words across forty-seven books, a decade and change of paragraphs nobody would reread now, all of it building the only thing that survived, a sense of when a sentence is true and when it’s faking. The practice is done. Delegating the execution now doesn’t rob me of a formation I already have. It spends it.

And here’s where the first unlock stops sounding like its own contradiction. Craft you don’t have — using a model to write fiction I never trained to write — looks like the opposite of earn it first. It isn’t, because the thing you have to earn was never the craft. It was the taste. Craft is how taste gets built, but once built, taste travels; judgment projects across gaps, which is most of what Standing on Surfaces That Move was about. The bedtime register I can recognize on contact but couldn’t produce, I learned from a lifetime of being read to and reading aloud, not from writing a hundred bedtime stories. So I bring earned taste to a craft I never practiced, and the model supplies the execution. That is not the kid’s move. The kid has no taste yet to bring, and no style, which is just taste transmitted through enough practice to leave a signature. A model can only project a signature that already exists: give it a developed style and it carries yours; give it a blank slate and it hands back the median. The maker with earned judgment gets amplified; the beginner gets overwritten, on time and instantly skippable.

So the rule that holds all of it together, without diminishing the craft that does the real work: you may delegate craft, whether you earned it or never had it, but only if you bring taste you paid for somewhere — and taste is the one thing the practice exists to build, and the one thing you must never hand off. Delegate the forging and you haven’t saved yourself the hard part. You’ve skipped the only part that was ever going to be you.

The “not really yours” part

There’s a sharper version of the objection: if the model wrote the prose, the book isn’t really yours. That one has a real answer, and it’s large enough to get its own essay , so I’ll keep it short here.

The work is mine at every layer where my choices bind. The premise. The character architecture. The ground rules of the world. The prose discipline, the named rules a project’s prose has to hold. The editorial pass. The decision to ship this version and not the seventeen others the model offered. The brand stance that refuses certain moves on principle. None of that is the model’s, and none of it is keystrokes. Where authorship actually lives once the typing stops being the work is a question worth taking apart on its own terms, and I do. For now it’s enough to say the obvious thing: the choosing is the authorship, and the choosing was mine.

The honest accounting

This essay is not claiming AI is neutral, not morally, not environmentally, not culturally. It isn’t. Taking these three unlocks means accepting a real bill. The models I use were trained on a vast amount of human work pulled in without consent and largely without payment, including creative work by people who’d perhaps object to exactly what I’m doing with it. I’m a party to one of the settlements that fight produced, so none of this is an abstraction to me. There’s the energy and water the systems burn. And there’s displacement landing right now on illustrators, writers, translators — people doing work I’m adjacent to, not safely distant from. None of it is small, and none of it is conveniently somebody else’s problem. My claim is only this: the bill is worth paying for some makers on some projects, with eyes open — and anyone who tells you it’s zero is selling something.

It also means the maker’s responsibility goes up, not down. Taking the unlocks comes with an obligation to spend them on work worth making. Producing slop with them is a choice. Shipping the seventeenth competent-stranger version of something we already have is a choice. The discipline the rest of these essays keep insisting on is exactly what makes the unlocks defensible in the first place. Without it, the cheat-code critics are simply right.

The alternative isn’t doing it by hand

The skeptic’s question — why AI at all — smuggles in an assumption: that the alternative is the same work, done without AI. For most of what I’m building, that’s not the alternative. Sojourn isn’t a kind of bedtime storytelling that happens to be easier with a model; without the model it doesn’t exist. A daily brief that weighs thirty-four feeds by perspective, run by one person, doesn’t exist without the breadth unlock. A continuously published, reader-responsive satirical paper run by one person doesn’t exist without the substrate. The choice was never between an AI-assisted version and a hand-made version of the same thing. The choice was between the thing existing and not existing.

So the arc that follows assumes you’ve made the call I made, that bringing the work into the world is worth more than the principled refusal that would keep it out. Plenty of reasonable people land the other way, and this isn’t an argument that everyone should take the offer. It’s an account of what becomes reachable if you do, and what it costs to do it without embarrassing yourself.

Unlocked, not cheated. The maker is still doing the only part that was ever hard. The unlock is just the part that didn’t used to be on the table.

Every senior engineer I know has felt this. The unit of an afternoon’s output shifted from functions written to decisions made and held. A senior’s edge over a junior used to compound through fluency. It now compounds through taste: the ability to look at a generated component and instantly see the four ways it will hurt the system in six months. A seasoned iOS engineer weighing a particular SwiftUI approach against a decade of watching the UI layer turn over: nibs and storyboards, the hard-coded-frame madness, Auto Layout constraints, and whatever comes next.

The “taste is the final human bottleneck” framing isn’t original. It’s been espoused by many in the AI community. The narrower argument I want to make is about who is feeling the move first, and what those of us on the frontlines might usefully say to the people about to feel it next. Software is the tip of the spear. The same shift in where the value lives is already reaching the creative arts, and the early indications are that the creative world is meeting it with a posture that won’t serve it well.

A field that was never stable

Speaking, I suspect, for most of us in tech: we take it seriously. Watching the shape of your craft change underneath you costs something, and the version of it we’re in right now is bigger than anything we’ve seen in decades. It may be existential. But it’s a more aggressive version of a feeling those of us in software have lived with continuously.

The surface keeps moving — yesterday's cutting edge, in a drawer.

Tools change. Frameworks die. Languages eclipse each other. The hottest thing in our dev stack last year is a maintenance burden this year, and a punchline next. Anyone who has been in this field for more than a few years has the same internal habit: assume the surface is about to move, and let the part you actually care about — the thinking, the architecture, the taste — live underneath the moving surface.

This is the part of the story to extend more broadly beyond tech, because I think it can be instructive. The creative world’s posture toward AI right now is largely rebellious, and the rebellion is understandable; the harms are concrete and the livelihoods at stake are not abstractions. Some of that rebellion is a justice claim, and waving it through as mere resistance to change would be wrong. These models were trained on an enormous amount of creative work taken without consent and largely without pay, and the people who made it are watching it get metabolized into the thing that now competes with them. That grievance is real, it’s distinct from “the craft is shifting,” and the structural point I’m about to make doesn’t dissolve it. The two just aren’t the same argument. I take the livelihood side of it up later in this series, in The Gate Was the Product and The Forge Was a Side Effect. Here I’m after something narrower and quieter — and it’s the part I can actually add something to.

The instability that has been the everyday weather of software is now coming for the rest of the creative arts. Resisting it as if it were an aberration, as if stable craft is the baseline and AI is the violation, does not prepare anyone for the shape of the next decade. What’s worth cultivating is the flexibility underneath, the kind that lets you keep being valuable when the surface moves, and then moves again next year, and again the year after that.

The same move, one industry over

Production is getting cheap in the creative arts, slower than inference in a source code editor but on the same trajectory. The first plausible draft — of an essay, a chapter, an illustration, a marketing campaign, a song demo — is approaching free. And the tools that produce it are landing on almost everyone now, inside and outside the industry. What is not getting cheap, and what shows no real sign of getting cheap, is the judgment about which plausible draft is worth keeping, which one will withstand a knowledgeable audience, and which one is built on something true.

The architectural change is the same as in software: production cheap, judgment expensive. The blunt version of the advice that follows is: find your taste, or be swept away.

What follows is the careful version, and it’s the interesting one; the blunt version, left alone, reads as callous.

The naive claim, and the truer inverse

The naive version of “find your taste” goes like this: the rare valuable person in the new world is the person who spans both disciplines: the engineer who writes well, the novelist who codes, the designer who can also illustrate, the producer who can also mix. AI rewards the polymath, and if you’re fortunate enough to be one, you’ll thrive in this new world.

That’s half right but it’s not the interesting half. Polymaths have always been valuable, and AI doesn’t change the basic shape of that advantage. The new thing is closer to the inverse. The irreducible human contribution to creative endeavors is taste: judgment, knowing-what’s-right, the ability to recognize when something works and when it’s three degrees off. Taste does not come from the AI model. Taste is what the person brings that the model can’t. Taste is experience across a life lived — the way relationships and hobbies and a profession, love and loss, scabs and stitches, compound into a person who knows what’s right.

What’s new is that AI lets taste project across gaps in craft the person doesn’t personally span. That is the whole game now: your judgment can reach into media your hands never learned. The engineer who couldn’t write the prose; the writer who couldn’t hold the system architecture; the designer who couldn’t sketch what was in their head; the composer who couldn’t engineer a finished mix. Previously each of those people was blocked by a missing half. Their taste was real and trapped, because shipping anything required mastering the half they didn’t have.

AI fills in enough of the missing half that the irreducible thing, taste, gets to project at full range.

This is a different claim from “the rare person spans both fields.” It’s the claim that you don’t have to span both fields anymore for your taste to express itself across them. The model is the bridge. Taste does the steering. Craft, traditionally the gatekeeper, becomes something more like terrain, still real, still consequential, but no longer the thing that decides whether you get to make the work at all.

The bottleneck used to be craft. Increasingly it’s taste.

One consequence worth addressing: the people who feel most threatened by AI tools and the people who feel most freed by them often have similar skill levels. What differs is whether their taste already exceeded their reach. If you had taste your craft couldn’t keep up with, AI is the most liberating piece of technology you’ll touch this decade, perhaps ever. If your value was in the craft itself, untethered from a strong opinion about what should be made, the ground is moving, and not in a fun disco way.

What taste actually looks like in the work

That’s the abstract claim. It has to cash out as something a person can actually do, or it’s a vibe. “Vibe” is the right word. In the AI conversation it’s become shorthand for accepting the model’s plausible output without scrutinizing it: vibe coding, vibe writing, the comfortable surface that gives way under any real weight. This argument about AI and creativity will fail the same way if it doesn’t translate into something you can practice.

The standard complaint about AI output is that it’s smooth and empty. Fluent on the surface, generic underneath, dissolving on contact with anyone who knows the subject. The instinct is to call this a soul problem, a creativity problem, some fundamental limit of the medium. It is none of those things. It’s a statistical problem with a specific shape.

A language model trained on a vast corpus, asked to describe a “deep red,” produces something near the center of how deep reds get described. The safe move runs toward the median, and the median runs away from specificity. A good writer might already have an eyebrow raised, knowing that specificity is key. The model is optimizing for plausibility across an enormous distribution of contexts it can’t see. It’s averaging. A knowing human, asked to name the deep red, says oxblood, or the red of the inside of a beet, or a gruesome murder scene. Each of those is wrong somewhere — out of register for some context the model is trying to be safe across — and that wrongness-somewhere is exactly what makes them right here. When you ask the AI to describe deep red, it’s never going to reach for Dexter.

So the discipline is this: in AI-assisted creative work, the human’s job is the relentless accumulation of true, scrutiny-surviving specifics that statistical smoothing can’t reach. Detail isn’t decoration on top of the work. Detail is the un-fake-able signature, the proof that a particular intelligence with particular knowledge was behind the thing.

A small example. A short story about an artisanal guitar builder will, by default, contain a workbench, some tools, a smell of wood and lacquer, and a craftsman who speaks in measured phrases. That’s the model running toward the median. The story becomes itself when the writer insists on the specific: a 1973 Martin D-28 with a hairline crack at the bridge plate that opened after the owner left it in a hot car in Phoenix; the smell of hide glue heating in a double boiler at the moment the kettle on the hot plate next to it starts whistling; the fact that a luthier will tap a top with a knuckle and tell you what wood it is before they look. Each of those is a thing a working luthier would nod at. None of them is in the median draft.

Specificity a working luthier would nod at.

That nod is the load-bearing thing. Specificity is what survives knowledgeable scrutiny. Anything smoothed toward the average tells a reader who knows the subject that no one who knew the subject was actually present.

An obvious objection, since I write this way myself: a model can be prompted to that bridge-plate crack and that double boiler. It can. Push it and it will hand you ten specifics, some of them good. But it doesn’t know which of the ten rings true and which is merely vivid, and it can’t, because that judgment is the part trained on a life rather than a corpus. Choosing — keeping the hide glue, cutting the detail that’s precise but wrong, knowing the luthier taps with a knuckle and not a fingernail — is the taste. The model can surface candidates all day. Which ones survive is still the human’s call.

The discipline scales inversely with how much the AI is doing. If you’re writing every word yourself, your specificity surfaces through the natural friction of composing. You’re picking each word because you mean that word. When the model is producing the first draft, the first draft contains none of your specifics. They have to be put in deliberately, paragraph by paragraph, with the patience of someone weeding a long row. The more production the AI handles, the more relentlessly the human has to supply the specificity the work hangs on.

This is what taste looks like when it’s actually working: an insistence, held over every paragraph and every decision, that the work hold up under the eyes of someone who knows.

The same discipline in software

In case the analogy from software to art seems too convenient, the same discipline is recognizable on the engineering side.

An AI-assisted engineer can ship a CRUD endpoint in twenty minutes. The endpoint will be plausible. It will compile, run, and pass the obvious tests. It will also, often, contain the median version of every choice: the median way to handle pagination, the median error response shape, the median naming convention, the median assumptions about the data. Each of those is fine in isolation and degrading at scale. A system built entirely from median choices is a system no one with strong opinions about it would have built. Yet it may work fine in some circumstances, that’s the rub.

The engineer’s job is increasingly the specific deviations from the median. The decision to make this endpoint idempotent because of how the production clients actually retry. The decision to surface this particular error class with a custom code, because three downstream services need to distinguish it from its neighbors. The decision to name this field after the domain language the team uses in standups, not the generic name the model offered. None of those decisions is the model’s. All of them are taste, applied as specificity, at the seams where the code meets the rest of the system.

The engineer who can hold a system in their head and notice every place the model produced a competent stranger’s version of a decision — that engineer is becoming more valuable, not less. The engineer who could only produce competent strangers’ versions is in trouble.

The honest caution

Soullessness is the first risk people name with AI-assisted work, and the specificity discipline addresses it directly: if you do the work of inserting the true, hard-won specifics, the output stops being generic. That one has a known answer.

The harder risk doesn’t, and it’s the part I’m least sure about.

Production efficiency tempts scale. If an essay used to take a week and now takes two days, the impulse is to write three essays instead of one. If a feature used to take a sprint and now takes two days, the impulse is to ship three features instead of one. The math is hard to argue with on the production side. The problem is that judgment doesn’t scale at the same rate.

An architect can hold one system in working memory. Two, with effort. Eight is comedy. A novelist can hold one long book in their head with the intimacy that lands the right specifics on the right page. Two, maybe. Five is fantasy. Taste is a slow-moving resource. It’s built through years of looking at the work, knowing what’s true in the domain, having opinions that survive contact with practitioners. It doesn’t speed up because the production tools sped up.

The honest version of the bottleneck-has-moved claim has to include this. The model works exactly as well as the human’s judgment scales, and human judgment is the irreducibly slow part. If we let production efficiency pull us into building more than judgment can steward, we’ll produce a great deal of work that looks fine and doesn’t hold up. It won’t fail loudly. It will fail in the way median work always fails: by not being remarkable, by not being remembered, by dissolving on contact with the people who would have cared.

That is the worst-case future of the cheap-production era: an enormous volume of competent strangers’ work, produced by people who had genuine taste and stretched it too thin.

Where this lands

Two things are happening at once.

The bottleneck moved. Craft was the gate for a long time, and the gate is opening. People whose taste exceeded their reach are getting to do work they could not have done before, and some of that work will be better than work produced by people whose craft never had to be steered by a strong opinion. That’s not a small change. It’s the most interesting thing happening in creative tools in most of our working lifetimes.

And: the discipline that meets the new bottleneck is the specific over the average. The taste that bridges craft gaps has to be paid out in detail-by-detail insistence, paragraph by paragraph, decision by decision, in the work itself. The model will not supply this. The model will produce the median, plausibly and quickly, and wait for the human to insist on something truer.

How far human judgment scales under the new conditions is the open question, and I don’t think it has a clean answer yet. What I know is that judgment doesn’t scale as fast as generation does, and that pretending otherwise will be the characteristic mistake of the next few years.

One last thing, addressed sideways to the creatives reading this who are still in rebellion. The instability you are feeling now — the sense that something you trained for your whole career has been moved without your consent — is, in software, the baseline weather. We don’t enjoy it either. But we have learned to put our defenses around the judgment underneath rather than the surface we happen to be standing on this year. That judgment is what stays useful as the surfaces keep moving.

Find your taste, then. Pay it out in specifics. Don’t let the cheap production talk you into making more than your judgment can hold. And remember that the version of the craft you’re being asked to defend right now is going to look outdated by the time you’ve finished defending it.

The bottleneck moved. The discipline didn’t.

And it is genuinely impressive, so sit in the applause for a second before the obvious thing arrives, which is: you’re not on the bicycle. You’re watching. The bicycle is riding itself across the stage and you’re across the room from it, clapping, a spectator at the experience you were ostensibly there to buy.

This isn’t entirely hypothetical. Engineers have actually built bicycles that balance and ride themselves, as a flex, and it’s a real piece of work. I’m not dunking on the engineering. I’m dunking on the pitch — the part that sells it as a product that frees you from something. A self-balancing bike that helps you stay up is one thing: an assist, training wheels for grown-ups, fine. A bike that rides itself, with nobody on it, is a different kind of object, and the kind is “a solution that ate its own problem on the way to the stage.” That part is hypothetical but worth imagining, because it’s less ridiculous than it sounds.

The category error

A bicycle has exactly one part you can’t remove, and it’s you. Pedaling isn’t the regrettable cost of getting somewhere — you are the thing getting somewhere. And yes, in a dense city a bike is also the best transport there is, faster and saner than a car, which is why half of Amsterdam commutes on two wheels. But even then you’re on it; the getting-there includes you. The self-riding bicycle removes the rider, and that breaks both versions at once. If you came for the ride — the wind, the effort, the small animal competence of staying upright and turning your own body into forward motion — it automates away the exact thing you came for. If you came to get somewhere, it delivers the bicycle to your destination and leaves you standing in the driveway. You were never the friction in the picture. You were the cargo.

It doesn’t save labor; it deletes the purpose. The bike gets there. You stayed home and got nothing. The benefit has no recipient.

The ride happened. He slept through it.

Means and ends

There’s a real distinction under the joke, and it makes the joke almost worth the absurdity, because it travels. Some activities are means: you do them to get an outcome, and if a machine can hand you the outcome and skip the doing, that’s pure gain. I do not want to wash dishes; I want clean dishes; the dishwasher is a small miracle and I have never once mourned the suds. Most of what we’re automating right now is means, and good riddance.

But some activities are ends: the doing is the outcome, the experience is the product, and there’s nothing left to hand you if you skip it. The bike ride. The meal you cooked for someone you love. The song you play yourself. The long conversation. The hike, as opposed to the arrival at the top of it. The self-riding bicycle is funny because it runs means-logic on an end. It treats riding a bike like dishwashing — a chore to be optimized — when riding a bike is the rarer thing where the chore was the point. Run that confusion at scale and you get a strange kind of progress: a world of flawlessly delivered outcomes that nobody is actually there for. I follow that nobody out to the empty grandstand it implies, from the watcher’s seat, in a coming essay, The Unwatchable Race.

The default direction

Here’s the part that’s less funny and more worth your attention. Almost nobody decides, in cold blood, to automate away a joy. You wouldn’t actually buy the self-riding bike. The joke only works because the error is so obvious when it’s a bicycle.

The trouble is that the error is almost never obvious. Delegation has a default direction — toward more of it, always — because each individual step is reasonable and convenient and saves a little something real. And ends rarely announce themselves as ends; they wear the costume of chores. Cooking looks like a chore until the night you realize you haven’t made anything with your hands for anyone in a year. Writing the thing yourself looks like a chore until the part of your mind that used to do it goes quiet and you notice the quiet. My handwriting sucks, it didn’t used to, and I hate that. The self-riding bicycle is loud and stupid and easy to refuse. The version that actually gets you is silent: a hundred small, sensible handoffs, not one of them obviously the one that mattered, until you look up and find you’ve optimized the doing out of a life and kept only the having.

That non-recognition is the real subject. It’s the thing I keep circling, in the abstract: we automate away things we didn’t know we were keeping. The self-riding bicycle is what that looks like drawn with a crayon.

Where “it relocates upward” runs out

Those of us who build with AI tools have a comforting line we tell ourselves: the effort doesn’t vanish, it relocates upward. Hand the grunt work to the machine and your attention moves to higher-value judgment. I believe that. I’ll argue it repeatedly, because for means it’s true and it’s liberating.

It also has a floor, and the floor is exactly at the self-riding bicycle. There are activities with no higher altitude to relocate to, because they were never a stepping-stone to anything above them. The bike ride doesn’t ascend into some purer essence of cycling when you automate the pedaling. It just stops happening. The joy doesn’t move up a level. It evaporates. Comparative advantage is a wonderful instrument and it bottoms out precisely on the things that were ends. And forgetting that is how careful, well-meaning people automate their own lives hollow with the best intentions.

The chairs are the wrong scare

You already know the cartoon version: the people in the floating chairs, fed through straws, gone soft, gliding through a world that does everything for them. It’s a tidy image and it lets us all off the hook, because nobody pictures themselves in the WALL-E chair. Of course you’d never let it come to that.

But the chair is the wrong thing to be afraid of. Nobody chose the chair. They each chose a hundred reasonable conveniences, every one defensible on the day, and the chair is just where a road paved entirely with reasonable conveniences happens to end. The dystopia isn’t a decision; it’s an accumulation. That’s scarier than laziness, because laziness you can resolve to fix, but accumulation only asks you to keep saying yes to things that are, one at a time, completely fine.

Not an argument against tools

I want to be careful here, because it would be easy to read all this as anti-tool, and it isn’t. I use the latest tools, every day, on purpose; most of the other essays here are about doing exactly that and why it’s worth the cost. I am not telling you to pedal to the office on principle, or to scrub by hand what a machine scrubs better. The line is means-versus-ends, not more-versus-less, and the whole skill is noticing which one you’re holding before you hand it off. Automate every means you can reach. Guard the ends, because the ends are the part of the life that was ever actually yours.

And which activities are ends is yours to decide, not mine. For somebody, cycling really is pure transport and the self-riding bike would be a gift, assuming it comes with a sidecar you can ride in. For somebody else the commute is the best twenty minutes of the day. I can’t hand you the list, and I wouldn’t if I could — that was the whole point of Becoming Gnarly . All I’m claiming is that the list exists, that it’s shorter than the marketing would like, and that it’s worth drawing on purpose, in your own hand, before the default direction draws it for you.

That’s one cut. The next essay, Standing on Surfaces That Move , makes the companion one — taste from craft, the judgment a machine can execute for but never supply — and the rest of these essays run on both.

The test

So here’s the test I’ve started using. It’s cheap and it works. When a tool offers to do something for you, ask whether you wanted the outcome or the doing. If it’s the outcome — clean dishes, a faster route across town, a draft of the boring memo — take the help and don’t look back. But if it’s the doing, if the doing was the thing you were actually there for, then what the tool is offering, very politely, is to be you at the one moment you most wanted to be yourself.

A bicycle that rides itself is not a better bicycle. It’s a very expensive way to not go on a bike ride.

The word for it is gnarly, which needs some defending, because to most people it lands as a surf-and-skate cartoon — a guy in board shorts, a turtle in sunglasses, a shaka hand sign. But the word has a second life the cartoon buried. A tree grows gnarled because it grew somewhere hard: wind, rock, exposure, a hundred seasons of not-quite-enough. The gnarl is the record of the struggle, and it’s also the reason the wood is strong and worth looking at. Smooth wood is young wood. The two meanings — gnarly the difficult, dangerous thing, and gnarled the weathered thing with character — are the same idea caught at opposite ends of a process. The difficulty going in is what produces the character coming out. That’s the whole thesis. The rest of this is me refusing to let it slip into being a fridge magnet.

The phrase itself, becoming gnarly, came out of a conversation about skateboarding with a close friend, and the grammar is the part I’d defend hardest. Not be gnarly. Not get gnarly. Becoming. You don’t arrive. In skateboarding there’s always a gnarlier line — a bigger gap, a taller wall, a scary trick that’s been made to look easy by everyone better than you. So the horizon keeps moving exactly as fast as you do. That sounds discouraging and is in fact the opposite. It means the thing can point a whole life in a direction without ever turning into a finish line you fail to cross. Chest-thumping needs an arrival. Becoming refuses one. It’s the most honest verb I know for a value, because it admits up front that you’ll die mid-sentence.

The receipts

A guy telling you to choose the hard path should have chosen a few, so here are mine. At seventeen I picked electrical engineering as my way into computers, and I picked it substantially because it was the hardest road to where I already knew I was going. I was right that it was the hardest — not kidding, I still occasionally have nightmares about exam failures, there were many. I’ve been collecting that kind of evidence ever since: forty-seven books written in the margins of a working life, a studio’s worth of products run by one person, and a stubborn habit of doing things myself well past the point where any efficient person would have hired it out. I love hand tools. I joke that maybe I was a grave robber in a past life because I love digging holes the hard way…I own a really nice shovel. I am DIY to a fault. And I mean to a fault, because that phrase is going to matter shortly.

I have the literal version too. I’ve skateboarded long enough to have collected the surgeries and the scars that come with it, and somewhere along the way I came to prefer a scar to a tattoo — not as a rule for anyone, just as my own taste. A tattoo is a mark you have applied to mean something. A scar is one that got applied to you, against your will, by something that cost you. You don’t pick the design, the fall does. The failure does. I have a skateboard with my name on it — a “bro model,” which in skating is the affectionate grade between amateur and pro, the board they make for the guy who is emphatically not pro but earned the ribbing of being pro-adjacent. That’s about the right altitude for all of this. Enough to have paid for the opinion. Not so much that I’d ever claim to have arrived. Becoming, not become or became.

My bro model — name on a deck (tiny on the turntable record), emphatically not pro.

Which is also, for the record, how I came to keep bees. In the things I do — skateboarding, mountain biking, ice hockey, the activities where gravity and the ground get to vote — injury is common enough that we have a name for what you take up while you heal: an injury hobby. A slam wrecked my shoulder badly enough to need surgery and a nine-month recovery, and nine months is a long time to sit still, so I went looking for somewhere to put the sidelined attention. Beekeeping sounded good, so I dove in. That’s the whole mechanism, and it’s gnarly all the way down: the hard thing hurts you, and instead of resting you go find another hard thing to be a beginner at. The bees in the backyard and the scars from the ramp are the same story — one’s just what I did while the other healed.

The injury hobby — where a blown-out shoulder and nine idle months went.

I left a part out of that story, and it matters more to me now than it did then. I kept at the bees for years — caught swarms by hand, ran as many as four hives at once, helped my dad with his. I ran different styles of hives (top bar and traditional Langstroth), learned how to keep bees organically, and read dozens of books and went to seminars and workshops. And most years I lost half of them anyway. Despite all the reading and the effort to keep the operation as low-impact as I knew how — no chemicals, no gadgets thrown at the problem, the bees mostly left to be bees — they died all the same, for reasons I still can’t fully name; the science only half understands them and my backyard was never a controlled study. After enough seasons the little ecology project I loved had come to feel like an expensive way to kill a few thousand bees a year, and I stepped back. The break has lasted longer than I meant it to. I might keep bees again — that’s the becoming, too, the part where the next hard thing turns out to be the one that already beat you. For now I’m a beekeeper on hiatus, which is a more honest line than the one I opened with.

There’s an upside to being a beginner that I didn’t expect and now hunt for on purpose. The obvious half is humility, getting ego checked, and that’s real. The half nobody mentions is that mastery levies a tax: once you’re good, the freshness drains out and the work turns to maintenance and small incremental gains. Starting over somewhere new refunds it — you get to be bad again, which is the only way to learn again in that steep, total way that only happens at the start. My friend Robert recently reminded me of the G.K. Chesterton quote, “If a thing is worth doing, it is worth doing badly.”

Of course, it’s easy to say, “go out and be bad and embarrass yourself in public.” The truth is there’s a price in the form of an ego hit most people won’t pay, especially in public. And I’m not pretending it’s easy to pay, but it’s easier if you know the trajectory along the way. At thirty I took up ice hockey: a kinda-sorta-near-pro-level skateboarder stumbling around the rink like he was trying to play basketball without knowing how to run, getting the puck poked off his stick by ten-year-olds. I loved it without reservation, exactly because I was terrible and there was that much open road ahead. I could feel that if I stuck with it, I’d get better. Becoming, again — same verb.

Not long ago a surgeon went into my knee to fish out a loose body — a chip off my own kneecap that, his best guess, had been drifting around in there for years before it finally started to bite. He was trying to work out what set it loose, so he asked what I get up to. The skateboarding earned a raised eyebrow. But it was the fixed-gear bike that stopped him — I was midway through explaining how a fixie fights you on a descent, no freewheel, your legs braking against the cranks the whole way down a hill, forced to turn with the wheel, when he just cut in: “Wait, why?” Not the knee mechanics. Why ride that bike at all. I sat there a second and said, “Good question.” I didn’t have the answer loaded, and the not-having-it nagged at me a lot longer than the knee did. Most of what follows is me, belatedly, trying to give him one.

The claim, plainly

Most things worth having are earned through real effort, and the effort isn’t a tax on the having; it’s frequently part of the having. The struggle to assemble a thing is a lot of why the thing ends up yours. This is old and unglamorous and some version of it came from your grandmother, or maybe her grandmother. The reason it’s worth saying again, now, is that we’re living through the largest effort-removal project in human history, and “no pain, no gain” is too dumb a slogan to survive contact with it. It has to get smarter, fast, or we’ll automate away things we didn’t know we were keeping.

Because here’s the part the gym-poster version gets wrong, and the part I most want you to take with you: not all friction is formative. Some struggle builds you and some struggle just costs you, and the entire skill — the actual discipline this whole publication is about — is telling the two apart. A hard route up a rock face makes you a climber. A bad line at the DMV makes you nothing but late. Treating all difficulty as virtuous is how you end up romanticizing inconvenience, and that isn’t a philosophy, it’s a martyr complex with a podcast. The point was never the suffering. The point is the specific, chosen difficulty that runs along the grain of what you’re trying to become. Everything else is just friction, and friction with no formation in it is precisely what tools are for removing.

Which is the honest reason I’m not against tools, and why the essays that follow this one spend most of their time using the most powerful ones we’ve ever seen. The line that matters isn’t human-versus-machine. It’s whether the difficulty you’re handing off was a means or an end. Delegate a means and you free yourself for the thing the means was always in service of. Delegate an end — automate the part that was the point — and you’ve deleted the thing while keeping its packaging. There is, I’m convinced, such a thing as a tool that performs your life on your behalf while you stand next to it and watch, and I’ll make that case at unreasonable length in the next essay. For now the principle is enough: keep the friction that forms you, shed the friction that only bills you, and accept that the work is knowing which is which — work that never fully resolves, which is the point.

The thing I can’t stop noticing

For most of history the central human difficulty was survival, and most of our species spent most of its effort on it. We have, for a meaningful slice of the world, largely solved that. And the people with the fewest imposed hardships turn out to be the ones reaching hardest for ever-more convenience — sanding away frictions so small we had to coin “first-world problems” to laugh at ourselves for minding them. I don’t think that’s decadence, exactly. I think it’s a muscle that doesn’t know what to do now that the thing it was built to push against is gone. We got the hardship removed and found we didn’t feel better, because some meaningful amount of who we are was assembled in the pushing. We were built to struggle.

Two things keep that observation from curdling into a lecture, and I hold to both hard.

First, this is about chosen difficulty, and I refuse to romanticize the imposed kind. Hardship that picks you — poverty, illness, grief, the struggles nobody volunteered for — is not a growth opportunity, and telling someone in the middle of one that the friction is good for them is obscene. The entire value here lives in the choosing. Strip out the choice and it’s just suffering, and I have nothing to sell about suffering. Can there be tremendous growth in suffering? Absolutely. But you don’t choose it.

Second, which difficulties are worth choosing is radically personal, and I’m not handing out a list. Using a dishwasher is not a moral failure. I’m not going to grind my own flour to prove a point, and the fact that I’ll happily spend years using hand tools to cut neighborhood mountain bike trails while ordering pizza delivered trailside so I don’t have to stop is not a contradiction — it’s the entire idea. You pick your hard things. The argument was never that you must suffer here rather than there. It’s only that there’s real value available in the struggle, that the value is easy to automate away by accident, and that it’s worth choosing some of it on purpose before the choosing gets done for you. Being yours to choose doesn’t make you right about the choice. It’s easy to christen a difficulty “formative” when you’re only flattering yourself. Your say-so isn’t the verdict; whether it actually built anything is, and that comes back from the work itself, or from the people who’d know — never from you alone. Incidentally, I have a friend who is grinding her own flour and having a wonderful time doing it.

Why these essays sit together

That’s the through-line, and it’s why the writing here belongs in one place even when it doesn’t look like it. The essays I’m publishing first happen to be about making things in the AI era, because that’s the enchanting easy road in front of me right now and the one where the lines are being redrawn fastest. But the subject was never AI. AI is just the most dramatic instance, this decade, of the only question I find genuinely interesting: what’s worth doing the hard way, and how do you tell. Underneath the software and the studio talk, every one of these is an argument about which struggles to keep.

Here’s the moment that taught me what the word means, better than any intimidating trick I’ve fought to land on a board. Years back, a few of us were skating a deep concrete bowl at a public park, all of us decades in, fluent the way you only get from a lifetime of slams. A greenway loops that park, close enough that you hear the runners. We’d been clocking the same woman for weeks. What she was doing barely counted as running — a pace most people would beat at a walk — but the effort was enormous. You could see what it cost her just to get one foot past the other, every step hard-won, her whole body in it. And she kept showing up. Day after day, in public, at the stage where it’s hardest and pays back the least. One afternoon I nodded at her and said, “…guys.” My friend Jim took a look and immediately said, “that’s so much gnarlier than anything we do.” Nobody laughed. It wasn’t a joke and everybody knew it. Eleven feet of concrete isn’t nothing, but it’s a drop I’ve had forty years to make easy; what she was doing took a courage none of us were spending. I don’t know if she ever clocked the ragtag crew of over-age skaters quietly in her corner, but I like to think being her cosmic fan club for a season counted for something.

I don’t expect to finish becoming gnarly. The process guarantees I won’t. The tree never stops being shaped by the weather. The line is always a little past where you are. You choose the next hard thing, you pay for it, you carry the mark, and you go looking for the one after that. The hard way isn’t always right. But often enough to organize a life around, it’s where the good stuff turns out to have been hiding.

So, welcome. Take the harder path when it’s the one worth taking. The rest of this is just me trying to work out which ones those are.