I.
A video blew up online recently. The influencer Cai A-Ga took his child to have a street artist draw a portrait, and while they waited he filmed it, telling the kid with a grin: "One of these costs two hundred, you know — I'll just have AI draw you one when we get home."
The backlash that followed isn't the point. The point is that the line was honest. He wasn't provoking anyone; he simply said out loud what a great many people actually think, the kind of thing few would say to your face.
The moment I saw the clip, I thought of 21 Lessons for the 21st Century, which I'd reread not long before. In it, Harari predicts that AI will produce a "useless class." When I first read that — there was no AI yet — I thought: how could that be, it's overblown, some distant and theatrical scenario.
Yet here in 2026, Cai A-Ga's line was dropped in ordinary conversation, to his own child, without a flicker of caution. A prophecy that ought to have provoked argument has become casual enough to pacify a kid — and that gap unsettles me more than the prophecy itself.
II.
Hidden in that line is an unspoken premise: I want a picture. Once the premise holds, the answer slides naturally toward — "as long as it looks good enough, who made it doesn't matter, and AI is faster and cheaper."
But that premise mistakes, from the very start, what the person actually wants.
Someone who thinks this way is mostly not looking for a work of art. He wants something to hang on the wall, to fill a blank space, something that looks decent enough. That need never required a painter — even without AI, he could buy a printed poster from IKEA. All AI has done is give "go to IKEA" a cheaper, more immediate version.
What's telling is the scale of the reaction the clip drew. The flood of comments and shares shows that many people, on seeing that line, intuitively knew something was off — even if they couldn't say why. What the street artist sits there making, and what the child pulls out of a computer back home, are two different things, even if the two pictures look alike. The intuition is right; it just hasn't been spelled out.
So the essence of the problem isn't that AI can replace human beings. It's that the person who said the line never wanted a painter to begin with — he didn't really want one before AI existed either.
This road never led to the artist, and never threatened the artist. The real question to ask is not this one.
III. Harari's Horse
But that road in section two — the one that doesn't lead to the artist — makes me think of another road that genuinely does.
The reason Cai A-Ga could say that line without any caution is a premise most people have quietly accepted: some of the things people do, a machine can do just as well, faster, and cheaper. In 21 Lessons, Harari pushes that premise to its limit and gives it a name — the "useless class."
It doesn't say that some people will lose their jobs. It says that an entire group will never find a next job to move into. When I first read it, I thought it overblown; now it lies open in a video made to soothe a child.
Every technological revolution has brought the same panic — the power loom, the automobile, the computer — each time someone cried that a whole cohort would be made obsolete, and each time jobs no one had foreseen sprang up. If this is the same script, then the shock is merely history repeating, and not worth writing about. What has to be tested is whether this time is actually different.
Taken all the way down, what holds up is not the blurry impression that "AI has replaced a lot of jobs" — that impression is too easily inflated into a sensational headline. What holds up is a harder-to-refute version: the problem isn't "AI took some jobs," it's "whether there is a whole group who will never find a next job to move into."
The way there were no more horse-carriages once the automobile arrived.
Harari's own metaphor is merciless: we may not be the drivers who retrain for new work — we may be the horses that get replaced.
The horse isn't unemployed. The horse has its entire role taken out of the economy, and no new job waits ahead of it. This is sharper than the four syllables of "useless class" — the question is no longer "will anyone lose a job," but "is there a role that gets removed wholesale."
IV. Translation: A Staircase Being Dismantled
Translation is the clearest place to test this. AI's translation is already good, so the linear inference runs like this: the machine first eats low-end document translation, then ordinary interpreting, until all that's left is the top-tier live interpreting — where a single wrong word at a heads-of-state summit sets off a crisis — held by humans. Along this line, humanity makes its last stand at the tip of the pyramid.
But go one layer deeper, and even the tip can't be held — not because AI has caught up to the top, but because the road to the top has been torn down.
No one begins standing at the summit interpreter's post. He climbs up through countless low- and mid-tier jobs: taking document translation to eat, accumulating real combat through ordinary business interpreting, getting it wrong, fixing it, getting it wrong again, fixing it again, training for ten years before he can even stand in that unforgiving room. Those low- and mid-tier jobs were never merely work — they are the staircase, the only passage by which a person can stay alive while training his way to the top.
What AI has eaten these past few years is precisely the lower flights of that staircase. And the waterline it has consumed won't stop at the middle; it climbs: the jobs that still feed a person this year are eaten next year; the position that counts as the top this year becomes next year's new middle, and goes on being swallowed. So the problem isn't "the top is narrow, so nobody wants it" — winner-take-all fields never lack for people crowding in; the narrower it is, the more people gamble on it. The problem is that even if they claw their way in, the staircase below has already been dismantled, and no one can climb ten years to the top while staying alive. The top empties not for lack of appeal, but because no one gets the chance to grow into the top anymore.
Acting works by the same logic. For an actor's craft to reach full maturity, he has to train from supporting roles, build a record, before he gets the chance to hold leading parts, before he has the name to choose his scripts and directors, and finally to grind his craft to its peak under a genuinely good script and a genuinely good director. But in an age when AI mass-produces images cheaply, the young can't find even the lowest roles, and the road upward is a non-starter.
What vanishes on the supply side isn't willingness — it's the path.
So the final possibility is this: it isn't that humanity is unwilling to have real people do the most important translation, or that humanity doesn't want the finest human actors — it's that we can no longer produce such people. Humanity is forced to accept all-AI translation and all-AI actors.
V. I Thought There Was a Clear Line Between Translation and Painting
I used to think translation and painting were two different things.
Translation is a tool. Its duty is to carry meaning cleanly across the river, to let both sides understand each other; as long as that function is served, whether a human performs it hardly matters. Reading Harari this time, I didn't care in the least who the translator was — and that is the norm for translation. The book is the destination, the translator the path, and swapping the person doesn't change where you arrive.
Painting looks different. What a buyer wants was never "a picture," but "the thing this particular painter made"; the person himself is what gets traded.
The one exception in translation that approaches painting is Lai Ming-chu and Haruki Murakami. Her feel for language, her choices, which shade of ambiguity she renders the original's ambiguity into — with Murakami these formed a pairing that can't be copied intact. But this is a coincidence bordering on the miraculous — ninety-nine percent of translation is like my reading of Harari: swap the path, and the destination is unchanged. Lai Ming-chu is the one percent.
A one-percent probability is an exception, and exceptions appear only in what is accidental and non-reproducible.
VI. Does Identity Matter?
If style itself can be learned — learned until the reader can't tell the difference — does "who made it" still matter?
Having thought it through, the answer is that it matters, but not for the reason I first assumed. Even if AI learns that style down to the last hair, that is only the shadow of the pairing, not the pairing itself.
The pairing of Lai Ming-chu and Murakami will end, but the way it ends will never be "replacement" — it's that two people must grow old and depart; fifty years on, whoever goes first, the pairing no longer exists, and no second, identical version will appear, because it was never a formula that could be re-run — it was the result of two specific people colliding at a specific moment in time.
Painting is the same, and purer. Two pictures nearly impossible to tell apart by technique — the instant one is authenticated as not by that specific person, its value collapses to near zero. The entire infrastructure of the art market — authentication, provenance, auction records — exists for one purpose: to put a price on "who made this."
So the only thing that truly separates "will be replaced" from "won't" is this: whether what you sell is a capability or an identity. Capability can be copied — no matter how high you stand, as long as the output can be cut off and sold on its own as a tool, the waterline catches you sooner or later, and that line can rise without limit. Identity isn't on that axis. It isn't a height; the waterline can rise and never reach it — not because it stands high enough, but because it simply isn't in the dimension that floods.
Translation sells capability, so even if the top survives, that's only an accountability premium — buying a living person who can be held responsible, to anchor the room where nothing may go wrong; no one cares about an interpreter's "style." Art sells identity. That is its only moat, and the watershed between it and every other craft.
VII. But Beneath the Moat, the Soil Is Washing Away
Put this way, it seems the artist has been handed a get-out-of-death card: identity can't be copied, the waterline can't reach it, safe.
This is a trap.
The artist himself stands on both axes at once. His identity can't be copied — that half is safe. But the entire formative process that grows this identity stands one hundred percent on the capability axis: he too climbed up from low-end jobs. That two-hundred-dollar street portrait, those mid-tier commissions taken to make a living, those years of jobbing no one remembers — all of it lives on the capability axis, and all of it is being eaten by the same waterline.
So that line in section one — "I'll use AI to draw it at home" — removes not art, but the very bottom rung of the capability axis. It is the same staircase as translation's dismantled one. What Cai A-Ga's line truly pulls out is the lower few rungs that let a young person live off drawing and hold on long enough to grow into "someone worth trading with."
The asset that is identity can stay immune forever, but the production line that manufactures the asset is breaking. Twenty years from now the artist won't vanish because identity has devalued — that's too far off. The real crisis is much nearer, and much more insidious: the asset is immune, and the production line snaps. The last cohort of artists able to grow up naturally may already be on the road, and the road behind them is flooding.
VIII. New Art Will Come; the Iron Law Won't Change
Writing to this point, I have to turn back and deal with the strongest rebuttal I set aside earlier — every technological revolution in history has produced jobs no one foresaw, so what makes this time an exception? Perhaps AI doesn't only tear down the old staircase; perhaps it simultaneously grows an entirely new art, one that could exist only after AI, and there are plenty of new positions there for new people.
This rebuttal is correct, and I don't intend to dodge it.
New art will certainly come. Photography wasn't a faster portrait; it made the "decisive moment" possible, something painting physically cannot seize. Film wasn't the moving photograph; montage made time itself an editable material. Each new medium's legitimacy came not from "doing it better," but from "doing something the old one couldn't." AI is likely the same, opening a space oil paint can't enter — a work that mutates continuously with the viewer and with time, with no final state, a generativity the human hand can never reach.
So the question shouldn't be "will new art appear" — that's nearly certain. The question is whether this new art escapes the sieve from before.
Imagine a fully autonomous creative subject made of AI. It roams the internet, is contaminated by everything it meets, and then grows its own works, with no human hand directing from behind. It's new enough, native enough. Now measure it once with the ruler from before: can it be traded?
The answer turns on one thing — whether the path of its contamination can be copied. If its state can be backed up, rolled back, duplicated to feed the same thing and get the same it, then no matter how new it is, it is infinitely reproducible, and scarcity drops to zero. It can be an important cultural phenomenon, a work written into the history of AI art, but it cannot sustain a market. It falls back onto the capability axis and is flooded.
The only way it survives is if the data-stream it ingested cannot itself be replayed — a unique trajectory of some real moment, occurring only once, that the universe cannot rewind and rerun. Only then is its history of contamination truly irreversible, and only then does it grow an anchor that can't be copied.
See this step clearly and one thing emerges: even the most native AI art can't escape the same sieve. What changes is the medium; what doesn't change is that what can be traded is always only the half anchored to the non-reproducible.
IX. The Last Anchor
So I have to revise what I said earlier. I kept calling art's moat "identity," but identity is not the bottom-most thing.
A specific person is an anchor not because of what he is called, but because his flesh irreversibly endures and ages in physical time, and that enduring cannot be replayed. Identity is the effect; irreversible time is the cause. The same logic in reverse also explains why AI's output is so cheap — not because it has no author, but because its time can roll back, can copy, can redo.
So the axis that truly can't be flooded was never "who made it," but "whether this stretch of time can be redone." Identity is merely a label that irreversible time leaves on human flesh.
This also makes clearer to me what I've been doing on canvas for twenty years. Multi-layer glazing is an anchor not because I decide not to scrape the glaze layers off — if it were only my decision, I could repaint one anytime. Its anchor lies in this: once the glaze dries it physically cannot return, regardless of whether I want it to. Time pressed irreversibly into the pigment, layer by layer — that process itself cannot be replayed.
X. The Horse, Asked a Different Way
Will the artist become that horse?
He will, but not in the way most people assume. Not the day the authentication system is forged or hacked — that's an arms race, and the market will counter with stronger provenance; the economic role of "who made it" won't disappear because of it.
What truly turns the horse into a horse is the demand side quietly redefining itself: one day, people simply stop caring whether the author is human. And before that day arrives, there is an earlier, quieter break point — once the staircase that feeds the young is cleared away, no one will climb to the position of being "worth caring about" at all.
The horse isn't saved by the waterline stopping at some line. It's that the line keeps rising, rising over its head. The only place the artist stands above water is the axis of irreversible time; but he must first survive long enough on the capability axis to climb there. And that stretch of road is being drained.
This isn't a question an artist can answer alone. It finally becomes a question put to all of human society: in a technological world where time can be rolled back, copied, redone, what thing's time is truly irreversible?
That thing is the last anchor.
