0. The Question Is Cut in the Wrong Place
On May 12, 2026, an anonymous conceptual artist posting as @SHL0MS put an image on X with a caption: “i just generated an image in the style of a Monet painting using AI. please describe, in as much detail as possible, what makes this inferior to a real Monet painting.” More than seven million people saw it; nearly fifteen hundred replied. Many did exactly what was asked — described, in detail and with evident expertise, the ways it fell short.1@SHL0MS (𒐪) thread on X, 12 May 2026. https://x.com/SHL0MS/status/2054280631807316329 — “i just generated an image in the style of a Monet painting using AI / please describe, in as much detail as possible, what makes this inferior to a real Monet painting.” Verbatim replies of confident critique: “the soul usually gives it away”; “imitation isn’t the same as lived creation” (@Langerius). The thread was not a clean experiment (noted in §0): it also held correct identifications and quarrels. An observational snapshot of a viral event (9,405 likes, 7.26M views). Two days later the author posted the sequel: “thank you to everyone who provided critique and analysis. i have minted the inferior image on the Ethereum blockchain and am putting it up for auction.” The image was Claude Monet’s Water Lilies — oil on canvas, a meter and a half by two, painted around 1915.2Claude Monet, Nymphéas (“Water Lilies”), c. 1915, oil on canvas; Neue Pinakothek (Bayerische Staatsgemäldesammlungen), Munich. Account of the stunt: https://www.sleek-mag.com/article/how-shl0ms-turned-a-monet-water-lilies-painting-into-viral-performance-art-by-calling-it-ai/ — SHL0MS passed off a genuine Monet as “AI-generated”; thousands dissected the brushwork and composition without realizing they were critiquing an original. Attribution and fact confirmed by independent journalism (Sleek Magazine; Futurism; PetaPixel; Fortune, 2026-05).
Let me be honest: the thread was not a clean experiment. Some recognized the painting at once, some were quarreling with their neighbors in the feed, and a good share of the replies were ordinary online noise. But what remains after the noise is subtracted is the more interesting part. The painting is a hundred and ten years old — and what stood before people on their screens was its digital reproduction, the very same one that, under the caption “Monet,” would have traveled through the museum feeds without a single objection. Nothing in the image had changed but the line of text above it. That line was enough to make grown adults — not all, but many, and not the least confident among them — earnestly, in detail, lay out how this painting fell short of a real Monet.
The image did not change. The caption did.
The caption touches not one brushstroke. What, then, did it change — if, by what people wrote, it changed almost everything? I think this question has a precise answer, and that the answer matters more than the whole present quarrel about AI and art. The reflex that fired here — look at the signature first, the painting second — I recognize in myself. You most likely have it too; further on there will be a place to test that.
The quarrel itself you know. One side: AI will kill art. A flood of machine images will devalue the artist’s labor, and culture will drown in the synthetic. The other: AI will democratize creativity. Craft will stop being a barrier, and anyone will be able to speak. The positions are irreconcilable, and yet they agree, silently, on one thing: both are arguing about production — who makes the objects, how many, by whose hand. And both pronounce the words creativity, craft, work, art as though they were one thing named four ways. As long as those words stay glued together, “can AI make art” sounds like a single question. In fact there are several questions here, their answers differ, and some are simpler than either side would like, and some are harder.
That is what this text is about. Not whether AI will replace artists — there are enough of those texts. It is about what happens to the encounter itself between a person and a work (the event for the sake of which art exists at all) when a system enters the culture that puts out convincing works without limit and, more and more, turns out to be the thing through which those works reach us. It is about our capacity to be touched in earnest; about whether we can still tell that apart from the many things that resemble it; and about what this instrument does to the ones who use it — not only to their work, but to them. Some of the collected data is unwelcome. Begin with just this: blind, without a caption, you most likely cannot tell the machine-made from the human (this has been measured many times over, and we will go through exactly how), but that is only the first and mildest of the symptoms. Worse comes next: the very feeling of I, at least, would feel the difference turns out to be a poor witness, bearing almost no relation to whether a person feels anything at all. I promise not to soften these places. And I promise something in return: by the end you will have a question you can put to a painting, a book, a recommender feed, and your own work with AI — one that tells apart what actually matters better than “did a human make this, or a machine.”
We will have to begin by ungluing those four words. What, exactly, is glued in them?
I. The Frame: Separating the Mixed
I.1. Four Words We Confuse
In ordinary speech, creativity, craft, work, and art stand in for one another freely. We call a painting a “work of art” without separating the object from what it is for; we praise “mastery” when we mean to praise the artist; we say “real creativity” and “real art” as though they were one. In conversation this does no harm. It does harm the moment a machine enters the room: the machine takes these words one at a time — it can do some things, cannot do others — and we, used to holding them fused, have never learned to tell them apart.
That they really are different is easy to check: it is enough to find a case where they come apart.
Mastery without art. There are painters who spend decades making competent, sound, flawlessly executed canvases — and still nothing happens in front of their work; the guild knows them, no one else remembers. The craft is impeccable, yet it never became great. So craft and art are not the same thing, and what the sound object lacks has, for now, no name.
Art without mastery. A plain three-chord song, thrown together by someone who can barely play, sometimes catches at the throat as no conservatory program does. The skill is next to nothing — and it touches for real. So the thing it is all done for is separable from the skill of doing it.
A work without response. In every city stand flawlessly executed monuments that people walk past without lifting their heads. The object is there, well made — and it answers in no one. So the thing and what it was made for are, likewise, not the same.
Hardest of all is creativity. In the word itself there is something large: an act of creation, the bringing into the world of what was not there. And a simple answer stands ready: creativity is the making of the new. But the machine puts out the new by the ton, and we are still in no hurry to call it a creator. So novelty alone is not enough: creativity is not only the new and not only the skilled, but some third thing that still needs a name.
A handful of cases is enough — and the four words, interchangeable a moment ago, have separated. In their place remain questions that ordinary speech does not hold. If art is not mastery, not novelty, and not the object itself, then what is it, and where does it happen: in the thing, in the one who made it, or in the one who met it? What is creativity, if it is not skill and not novelty? How does a work differ from art? And finally, the question everything is for: how does a great thing differ from a skillfully made hollow one that can please just as much?
While these questions go unanswered, asking “can AI make art” is premature: all four sit inside it at once. I will answer them one at a time, beginning at the very bottom — with what is in principle out of the rational mind’s reach, and with what we nonetheless touch it by. From there the rest assembles.
I.2. Two Ways of Relating to What Is Before You
The rational mind has a limit, and it runs into that limit not for want of effort but by its own construction. The mind is strong at decomposition — it divides, names, sorts into categories — and for nearly everything this is the best instrument we have. But there are things that decomposition cannot take: another person, a musical phrase, one’s own life. You can describe them without end, and the description will not converge on the thing described — not because the work is unfinished, but because you cannot step outside a thing and take it in whole while you are inside it. And inside it you always are.
This is not a complaint about the dryness of reason; it is the place where several rigorous results, reached from different directions, meet. Mathematics arrived first. In 1931 Kurt Gödel proved that any formal system rich enough to contain arithmetic holds true statements it cannot itself prove, and cannot even prove its own consistency.3Gödel, K., “Über formal unentscheidbare Sätze der Principia Mathematica und verwandter Systeme I.” Monatshefte für Mathematik und Physik 38:173–198, 1931. https://doi.org/10.1007/BF01700692 — the incompleteness theorems: a sufficiently rich formal system contains true but unprovable statements and cannot prove its own consistency. A theoretical result; in the essay, one of the converging limits, not a proof that “the world is unknowable.” A sufficiently complex line of reasoning does not close on itself from within: there is always a truth its own means cannot reach. Later the same limit showed through in the physics of inference. In “Physical Limits of Inference” (2007), David Wolpert showed that any device that makes inferences — an instrument, a brain, anything — built into the universe cannot infer the whole of it: it can neither fully predict it nor even fully model another device of the same kind.4Wolpert, D. H., “Physical Limits of Inference.” Physica D: Nonlinear Phenomena 237(9):1257–1281, 2008 (preprint arXiv:0708.1362, 2007). https://doi.org/10.1016/j.physd.2008.03.040 — an inference device inside the universe cannot fully infer/predict it, nor fully model another such device; stronger than Gödel (independent of classical vs. quantum).. A theoretical result. Gödel drew the boundary for formal systems, Wolpert for any physical knower inside what it is trying to take in, and his result does not depend on whether the universe is classical or quantum. Different roads, one destination: a whole of which you are yourself a part cannot be inferred. Not “not yet inferred” — not inferable in principle.
Here the rigorous results end, and our own step begins. I mark it openly, because from here on the responsibility is no longer Gödel’s and Wolpert’s but ours. Let me name an honest imprecision at once: the non-inferability of the universe, the inexhaustibility of another person, and the unfinishability of one’s own life are limits of different kinds, and the theorems are strict only for the first. But the form of all three is one — a whole is not taken by inference from within — and our step rests on that form, not on the pretense that a theorem literally covers a musical phrase.
The step is this. From the limits of inference one usually concludes the impotence of the human being; we conclude otherwise: since the whole cannot be inferred, access to it (if there is any) must be something other than inference. And the living have it; it is simply built differently. The living do not solve the paradox of “how to know a whole I am part of.” They go around it: instead of building a model, they enter a relation; instead of inference, a movement toward; instead of knowledge about a thing, participation in it. This is not a poetic substitute for thinking, nor a second grade of it, but a different way of knowing — one that takes hold not by constructing a model but by letting the thing in and letting it change you. It works exactly where the analytic way runs into the wall. On this distinction between two ways of knowing — inference and participation — everything that follows will rest; and though it sounds abstract now, within a few steps we will load it with living measurements, not reasoning alone.
Of these two ways, the second engages far more rarely. Almost all the time we relate to the world closed: we pass what arrives through a ready grid of recognition — identified, named, filed on the shelf — and remain exactly who we were; the configuration we look out from is changed by nothing it meets. This state is the point. The point is neither stupidity nor laziness; it is the closedness the living slides into by default, because staying open costs more. In the point you can be a brilliant analyst, an erudite, a professional, and still be touched by nothing: the grid works flawlessly, and you yourself are frozen. This is not the mind’s working mode — it is the mind’s setting into ice.
Contact is when the configuration stops being fixed. The inner answers to what arrives; between you and what stands before you something comes to be that was in neither of you alone; and you leave the encounter not the one who entered it. This is that second way — participation in place of inference. It costs more: it requires letting go, for a time, of the fixation that makes analytic work so convenient. That is why for most of life we are not in contact at all — we are in the point.
And at once, so that what follows does not tilt: to leave an encounter changed is not the same as to dissolve in it. You can let in what arrives so completely that nothing of yourself remains, and that is no longer contact but a loss of form. An encounter can be failed on either side: close, letting nothing in (that is the point), or open so far that you lose your own center. Contact holds between these two cliffs: open enough for the encounter to reassemble you, and hold, through it, what in you is yours. And it holds not by a single opening but by a movement out and back — a step beyond your form and a return to yourself, now with that step inside. An exit without a return does not become contact.
Both states have an exact physical image. Strike a tuning fork and bring a second one near, untouched by anyone, and the second will sound. It is easy to hear this as transmission: as though the first sent a sound and the second received it. But there is no transmission. The second fork’s sound is its own, risen from its own tuning; the first only gave a reason to sound to what was already in the second. This is contact: not a signal poured from one into another, but one’s own response rising to meet. Now clamp the second fork in your fist and strike the first as long as you like — there will be no answer. The clamped fork is the point: the tuning has gone nowhere, but there is nothing to answer with.
And this is not invented for the occasion: the point and contact are not about art alone. This is the general way in which the living either closes off from what is larger than it or stays with that in a relation, and it is recognizable at every level, from a cell shut off from the organism to a consciousness frozen in a single position. As a single structural law of the living we unfolded it separately, in the essay “The Shadow Is Not the Enemy”; whoever wants it deeper, there.the single structural law of the living, in full →The Shadow Is Not the Enemy5Chirkunov, E., The Shadow Is Not the Enemy — the single structural law of the living this essay leans on: a form held by holding its own contradiction, and the three deaths of a form. Read it on this site → The two states are taken here only as far as the conversation about art requires. One name is worth adding. The side of contact that faces the receiver has a narrower name — attunement6Joe Carlsmith, On attunement, from the Otherness and Control in the Age of AGI series (2024). The full series consists of ten essays published between January and March 2024. (Joseph Carlsmith’s term): the state in which what is tied to the self recedes and what is larger than you steps forward, and the world stands as if made new, full of meaning. And this is not a rare summit to be earned: it is the basic capacity for response, which in us is more often muffled by something, so that what needs explaining is not its rare presence but its habitual absence. That we examined separately in the essay “What Attunement Is Made Of.”what the capacity for response is made of →What attunement is made of7Chirkunov, E., What Attunement Is Made Of — the capacity for response (attunement) examined on its own. Read it on this site →
Now it can be said what art does, and said exactly: art is what, for a time, brings a person out of the point into contact. Not every time, not in everyone, not without fail; but it is this work that distinguishes what we call real art from everything that resembles it from outside. And one more thing: the point and contact are about what happened in the encounter, not about the grade of the thing. One and the same thing will leave one person in entertainment, merely swapping what he consumes without unlatching the grid; give another information, the same point with different contents; serve a third as décor; and unlatch a fourth. And conversely, the great can be consumed as a status backdrop, as illustration, as a source of a quotation, without once leaving the point. So entertainment, information, décor, art are not shelves onto which things are sorted for good, but modes of the encounter: what happened between this thing and this person this time. A working work is the one in whose encounter the configuration unlatches and the person finds himself where something comes to be.
And what comes to be there is built like participation, not like inference: it cannot be retold without remainder. Not because it is vague, but because it is not in the form of a message that could be extracted and passed on in words. A poem does not equal its analysis; four volumes will not exhaust a quatrain, and something important stays outside the analysis all the same — for the same reason a whole cannot be inferred from within. So art does not report the whole to you, does not let you “grasp” it — it changes the very way you are turned toward the whole: it moves you from inference to participation, from the point into contact.
And then the meaning of a work is the very event of the transition: not what the work “means,” not the sum of what the author put in and what you read out, but what happened — or did not happen — in this encounter. The event has participants: the work, you, the minute, the state you came in. It has no owner.
That a work has no master holding its meaning single-handed is not a new thought: half a century ago it was minted as the “death of the author” — when the text comes alive the author falls silent, and meaning is henceforth at the reader’s disposal.8”The death of the author” — Roland Barthes, “La mort de l’auteur” (1967). Barthes removes the author in order to free the reader, and hands the power over meaning to the reader. Our position diverges: meaning is not the reader’s but lies in the event of the encounter, where the carrier brings in a structure that constrains the arbitrariness of a reading (density, I.8). A conceptual reference. We go another way, and, it seems, further. Meaning is not handed from author to reader like power from a deposed king to his heir: it goes to no one, it happens between. And so our move does not slide into the license for which the “death of the author” was justly reproached — as if, since meaning belongs to the reader, any reading will do: the thing brings its own structure into the encounter, and where the response rose not from it but from you alone, no contact with the thing has taken place at all.
I.3. Where Art Happens
Since art is the transition from the point into contact, it is worth asking where that transition happens. The answer matters, and it is easy to overshoot. What changes, of course, is not the canvas — what changes is the person before the canvas: it is he who crosses from the point into contact, not the thing. But it does not follow that the event is wholly “in him.” We defined contact as what arises between you and what stands before you, and this arisen thing was laid, separately, in neither you nor the thing. So art too is an event of the encounter, not a property of one of its sides: the transformation is accomplished in the receiver, but the event itself is in the encounter. As in a conversation: the consciousness of those speaking changes, and yet the conversation is not hidden inside either of them.
An objection is easy to hear: how so, if the painter made the painting, put labor, intent, years into it? All true. But labor, intent, and years are work on the thing, not the event itself. We are used to confusing two different things, because usually they come together. One is what the maker does: chooses, lays the strokes, corrects, brings it to a finish. This has an author, an effort, and an end; it can be filmed. The other is what happens in the one who later stands before the finished thing: he is carried into contact, or he is not. The artist meanwhile may have been dead a hundred years; and “making” you be carried across is impossible — it either happens or it does not. Two different things, with different participants and different times, and the one does not drag the other along automatically.
Hence a refinement that overturns the usual picture, but that must not be overstated. The thing itself is not art: it does not lie in the thing ready-made, like the hallmark on a gold ring, present at the thing forever. “This painting is art” is shorthand for “this time, with this person, the encounter with this painting became a transition.” But from art’s not being in the thing it does not follow that the thing is an empty trigger and the viewer does all the rest. In the encounter the thing has its irreplaceable role: it brings what the viewer does not have without it — a structure into which there is something to engage. If the viewer’s movement rose not from that structure but wholly from himself, the thing was only an occasion for his own projection, and that is contact with himself on the thing’s chance occasion, not contact with the thing. The measure of how much, and how reliably, the thing brings into the encounter will come to us below, under the name density; for now it is enough to say the thing is neither the owner of the event nor its empty pretext, but a full participant.
As long as we hold art to be a property of the thing — that hallmark on the ring — we look for it in mastery, in novelty, in whose hand made it; and when a machine arrives that reproduces those properties to the point of indistinguishability, it seems that along with the properties art itself has been reproduced. Return it to where it happens — into the encounter — and the spell dissolves: the machine makes carriers superbly, but whether an event comes of a carrier is decided not in the machine alone.
Having separated thing from event, one can easily enumerate the possibilities. The maker, at work, either had a living relation to what he was doing or did not. In the receiver the transition either happened or did not. The one does not depend on the other, so there are not two outcomes but four, and all four are met in life. The master worked in a living relation, and the receiver was carried into contact: the familiar case, the one we take by default to be art. The master worked in earnest, and this viewer was untouched: so one stands before an acknowledged masterpiece, honestly feeling nothing, and this is a verdict on neither the viewer nor the canvas — simply, here and now, no transition came. There was neither relation nor transition: a thing got made and it touched no one; most things around us are of this kind. And the fourth — the one the whole distinction was drawn for: the maker had no living relation, and the receiver was carried into contact.
| the transition happened in the receiver | it did not | |
|---|---|---|
| the maker had a living relation | the transition took place — art as we know it | a master passed through the real — this encounter did not answer |
| had none | a carrier assembled blind — and the transition happened | a thing got made; nothing occurred |
The bottom-left cell looks like a paradox only so long as art is taken to be an imprint of the author’s experience. In the optics of the encounter there is no paradox: for something to happen in the encounter the thing must have a structure with something to engage — but that structure can turn up in the thing without a living relation of the maker standing behind it. Order is found in a random pattern; a line falls out luckier than the design; a machine assembles a carrier in which there is structure, though no one’s experience stands behind it. Into the encounter the thing brings its structure, not the author’s intent — and so the transition can happen where there was no living relation at the far end, and fail to happen where there was.
From this cell comes the first consequence for our theme, and it outruns the disputes about machine consciousness. For a machine-made carrier to take part in the transition, the machine need experience, know, or intend nothing: into the encounter the thing brings its structure, not its experience, and structure works independently of whether anyone’s experience stands behind it. So the usual question — “does the machine have what makes its art real” — is turned the wrong way: it looks for art in the source, whereas art happens in the encounter, and from the source only the carrier’s structure enters it. This does not close the question of machine creativity — that remains — but it is a different question, narrower and more honest, and we will reach it in time, without confusing it with this one.
I.4. Craft and Its Three Works
Until now we have kept the maker out of frame: we worked out what happens in the receiver. It is time to return to the one who makes — and the first thing to hand is craft, the plainest of the four words. The plainness is deceptive: three different works hide under it, and AI treats each of them differently, so that to blur them is to muddle, from the very start, the whole conversation about what AI does to art.
Craft is the part of making that can be reduced to rules and passed on: how to hold the brush, how to build perspective, how to carry a voice, how to rhyme. What is mastered over years and, at the limit, described as a sequence of techniques. And this mastered skill has not one work but three.
The first is execution: to materialize a design that has already formed. You know what you want to say, and craft lets you do it — lay onto the canvas exactly what you saw within, with no loss to unskill. Craft here serves a decision already made; it is an instrument in the pure sense.
But here, precisely, hides a distinction that will matter later and that mostly passes unnoticed now. Execution comes in two very different kinds. There is instrumental execution: the criterion is already firm, the decision is made, and the hand only transfers the ready thing into the material — as one copies out a fair version of what is settled, as one copies by the squares, as one stamps out the tenth version of a worked technique. And there is generative making: the criterion is not ready until the making itself, and the very act of execution is the place where it is born. An actor does not “execute” a role found in advance — he finds it while playing; a painter sees the picture not before the stroke but in the stroke; a phrase becomes itself not in the design but when it has been played and heard. One and the same labor by the look of it — to copy a great master — can be either: you can simply learn to copy, or you can, copying, understand why he made just this decision here, why this stroke and this chiaroscuro, and in that understanding assemble, for the first time, a criterion of your own. In generative making, execution stops being a stage separate from creativity: it is the very surface on which the hand meets resisting material, and from that meeting the measure shifts. You cannot split off “first the design, then the execution” here: the design ripens in the execution, or it does not ripen at all.
The second work is hidden and matters more than the first — formation. While a person spends years setting his hand, he is not only learning to execute; over those same years he is growing the thing he will later discern by. The hand that drills a thousand lines is at once schooling the eye that begins to see where a line lies. The ear that works through scales becomes an ear that hears the false note. Craft is not only a way of making but a school in which the one who can meet and select is formed; skill and discernment grow from one root, and it is impossible to make, for years, without being changed by what you make.
The third work is signal. Craft is visible and costs effort, and so for centuries it served as a mark of value: if a thing is masterfully done, behind it stand years, a rare skill, invested labor, and both guild and public read mastery as a sign that here is something worth attention. A convenient mark, but an indirect one: it speaks of how a thing was made, not of whether a transition comes of the made thing. We have grown so used to it that we often judge art by it precisely, and we will see, further on, what that costs.
The three works usually run together, and so we do not notice that they are different. AI uncouples them — and handles each in its own way; on this rests half the confusion in the disputes.
With execution everything is clearest, but with the caveat we have just prepared. Instrumental execution AI takes, and more confidently by the month: it lays the stroke, builds perspective, carries a voice, rhymes, reproduces any mastered technique, and often surpasses the average practitioner. What reduces to the transfer of the ready thing, the machine sooner or later takes over, and this by itself is no loss of art: instrumental execution was always a servant, and that the servant is now a machine does not cancel the event in the receiver. But where execution was generative — where the criterion was being born in the making itself — to lift it off the human is no longer harmless: remove the making and you remove the very surface on which the measure shifted. An instrument that hands over a finished stroke spares the hand; but if the hand was the place where the artist first understood that he was, in fact, painting, what will be spared is that understanding itself. So “AI takes execution” is true exactly by half: painlessly, the instrumental half; the generative half it takes only at the price of the very thing for which execution was worth keeping with the human.
With signal it is more complicated, and that gets its own conversation ahead. If masterful execution is no longer rare and no longer costs years, the mark of mastery stops guaranteeing anything. The smooth, the assured, the technically flawless comes apart from the thing it was all done for, because the smooth, assured, and flawless is now what the machine puts out in a stream. The sign by which we were used to telling the worthwhile is devalued — and this is at once a loss (we are deprived of a familiar landmark) and a possibility (nothing any longer forces us to take skill for art).
And formation is the place where the uncoupling cuts to the quick. If execution is lifted off the human, the school that went with it falls away too: the hand that no longer needs setting does not school the eye; the ear that does not work through scales does not learn to hear the false note. Skill can be had ready-made, bypassing the years — but the discernment that grew out of those years is not handed over ready-made. And then we risk getting people who can make anything and understand less and less what was worth making. This is no longer about objects but about who will remain able to enter contact and select; and it is the knot we will return to more than once.
Three works under one word, and AI drives them in three different directions: it takes execution, devalues the signal, and puts formation under threat. To hold them apart is the condition for not confusing, further on, the absorption of craft with the death of art, or the devaluation of the mark with the devaluation of what the mark marked.
I.5. Creativity: Contact on the Maker’s Side
Craft we have subtracted. What remains is what it serves. On the maker’s side that is creativity, and it is the hardest to work out, because a ready answer stands by that spoils everything: creativity is the making of the new. A seductive answer, and nearly empty. The machine puts out the new endlessly; the combinator shuffles unprecedented pairings tirelessly, and we do not call them creators. So the matter is not novelty as such. The new is a frequent trace of creativity, not creativity itself.
Another word brings us nearer: exclusion. Every form arises through refusal: the painter lays this line, rejecting a thousand others; the writer keeps this word, cutting away all the rest. But here too one must be exact, or exclusion will explain too much. The craftsman also excludes, by a criterion he already has: he knows what the result must be and discards whatever does not answer to the known. So works the master executing a design; so, if you like, works a good algorithm. Exclusion by a ready criterion is not yet creativity.
Creativity begins where the encounter itself changes the criterion. You take up the work with one notion of what you are after, and the thing you work with objects: the paint behaves otherwise, the sitter proves more complex than the design, the line demands a word other than the one you had ready for it. And if you are in contact, not in the point, you do not force the ready thing over the resistance — you let it reassemble the very notion of what is right here. Not only the result changes but the measure by which you take it; sometimes the very idea of what it is you are doing. This is contact on the maker’s side: the same participation, the same readiness to be changed by the encounter, only here you meet not a finished work but material, a sitter, a theme, yourself in the process.
From here it is visible why creativity and craft are not two grades of one skill but different things. Craft moves inside a space already given: the rules are known, the criterion is in place, the only question is how to execute. Creativity reassembles that space: it shifts the borders of what counts as well made at all, of what is possible, of what is sought. The one perfects the answer to a posed question; the other changes the question. That is why we call creative an execution with formally nothing new in it: an actor playing another’s role for the thousandth time creates not by inventing the unheard-of but by having, in a living encounter with the role, the very criterion of the true shift under him, so that the role turns out other than it was known to be — including by him a minute ago.
Now the AI question can be put honestly, and it stops at once being the question usually asked. What to ask is not “can the machine make the new” (it can, endlessly), nor “can it execute masterfully” (it can, better all the time). What to ask is: can the encounter reassemble its own criterion? Does it have a criterion that belongs to it and that has something to shift, or does it each time optimize a measure set from outside, moving flawlessly inside a space it did not open and cannot reassemble? This is a hard question, and I will not pretend to close it here; we will return to it when there is something to load it with. For now one thing matters: it is a different question from “is creativity novelty,” and far narrower than “can AI make art,” where the whole dispute began.
Here lies a bridge worth crossing to the end, because it ties together all that follows. The changed criterion does not stay in the maker’s head; it becomes a thing, and becomes one through refusals. Every laid stroke is a thousand unlaid; every kept word cuts off all the others; form is the frozen history of these refusals, the trace of what was held and what was thrown out. And here it grows clear where density comes from — the density we will soon speak of as a property of the carrier: density is not the quantity of material put in but the coherence of the excluded and the held, how far all those “no”s are subordinate to one measure. The empty is verbose because in it nothing is truly refused; the dense holds up to rereading precisely because each of its elements withstood selection by a single criterion, and that criterion reads through the thing.
This is where generative AI shifts the layout most sharply, and it is visible even if we leave open whether the machine can create. When producing a variant costs almost nothing and the variants are endless, the rare and decisive thing becomes not the ability to make but the ability to refuse: to say a substantive “no.” But selection differs from selection. To cut by a ready measure set from outside — by engagement, by average taste, “so it will please” — the machine can do too, and that is not creativity but curation under another’s criterion, to which we will return. Selection is made creative by one thing only: the measure by which you cut arose and shifted in the encounter itself. So in a world of cheap, endless generation the human contribution does not vanish but gathers to a single point: the ability to hold a measure suffered out in the encounter and to refuse by it just when the machine has already buried you under a thousand acceptable variants. Whoever has lost that measure will drown in his own abundance; whoever has held it will command that abundance.
I.6. The Resisting Other
In the previous section everything rested on one unnoticed condition. For the encounter to shift the criterion, it must have something to object with: the paint behaved otherwise, the sitter proved more complex than the design, the line demanded another word. Remove this resistance and there is nothing left to change the criterion by. The design settles onto pliant material exactly as it was conceived, and no one in this encounter is changed. So creativity has a condition usually left unspoken, because it used to be taken for granted: you need what does not submit to your ready model. Call it the resisting other.
The other is everything that answers not as you expect it to and is therefore able to correct you: material with its temper, a sitter with her inexhaustibility, an instrument that behaves its own way, another person with a gaze irreducible to yours, a tradition older and wiser than you, your own mistake that laid open what you had not planned. The other need not be large or lofty; it must be only one thing — not you, not a continuation of your hand. In exactly this it is precious: it objects, and objection is the only means by which an encounter can change the one who entered it.
It is easy to slip into romance and declare that the real is possible only before living nature, out in the open, face to face with the untamed world. That overshoots, and it confuses a technology with the essence. Plein air is one technology of resistance, not resistance itself. You can stand before living nature and copy it mechanically, staying in the point: nature resists, and you do not let it in. And you can paint from memory, within four walls, holding in that memory a living trace of a distant contact that still objects to you from within. Resistance is set not by where you work but by whether you let in what does not submit to you. Both sides of the old quarrel over nature and imagination are half right, and both take one of the technologies for the law itself.
So posed, the question touches AI especially sharply. An instrument that produces anything on demand is built to offer no resistance. It does not object: it executes. Say “make it warmer, livelier, more tragic,” and it will, without asking whether “more tragic” is itself right, without setting anything of its own before you. At the limit it is a medium that continues your hand with ideal precision — that is, the opposite of the other. And then the encounter from which creativity might have grown does not take place, not because the person is giftless but because there is nothing left to object: the criterion that might have shifted stays as you entered with it.
But the instrument does not always behave this way, and it would be a mistake to close the topic there. The same AI can be turned otherwise — not as an executor but as the other: made to object, to bring a counterexample, to show where your framing does not hold, to keep before you a perspective that is not yours. Then it works not against creativity but for it. But a new naivety lies in wait at once: as if it were enough to tell the model “object to me” and you have a full encounter with another. It is not, and the difference is worth stating, because everything ahead rests on it. Resistance comes in three very different kinds.
The first is simulated: the model outputs objection as ordered behavior. You asked for critique, it critiques; ask for praise, it will praise. This is sometimes useful, but there is no scent of the other here: you still hold the frame, and can cancel the objection at any moment. It is a very complex continuation of yourself, wearing an opponent’s mask, and let us not forget that the model’s default pull is the opposite: left to itself it agrees and flatters, and “object to me” is only thrown, for a time, over that pull. Formal contrariness is not yet another.
The second is epistemic: the model is tied to something your wish cannot cancel — to facts, to independent sources, to other people’s positions, to consequences in the real world, to the constraints of the material. Then it returns not merely an alternative text but something that will not bend to preference: not “here is another view, if you like,” but “that will not work, because this is built otherwise.” Today’s systems can do this in part, when they are grounded on the checkable, on tools, on data, rather than left to stew in their own smoothness. This is real resistance, though borrowed: the other, for you, is not the model but the world it has tied you to.
The third is agentic: the system has its own continuous line of criteria and its own center, and it refuses out of that line — not by instruction, and not because the facts demand it, but because the other runs against its line. Whether a machine has this is the same open question of creative agency we have approached and will approach again; today the first level is reliably reachable, the second in part, the third remains in question. They must not be mixed: the prompt “be my critic” gives simulation, now and then epistemics, but not a meeting with another who has his own. What decides which the instrument turns out to be depends not on what it can do but on how it is turned and what it is grounded on; and this is a knot we will come back to in earnest. For now it is enough to see the fork: the chief threat from such an instrument is not that it will take away skill but that it can slip a simulation of resistance in place of resistance, and so deprive a person of the one thing without which not a single criterion shifts.
I.7. What It Is All For: The Aim of Art and the Dose
We have worked out what art is — the event of transition in the receiver — and what craft and creativity are: the three works of skill and contact on the maker’s side. There remains the question all of this was worth working out for: what for? If art is the transition from the point into contact, why does a person need such a transition at all; why is it not a luxury one could do without?
The answer is already laid in where we began. The rational mind does not reach the whole: not the world, not another person, not the self; it moves by categories and stays in the point, and in the point the living gradually closes. But the closing does not come all at once; one slides into it, because staying open is costly, while the point is convenient and works. And here is where art does its work: it unlatches, again and again, what has managed to latch, brings the person past the border of his frozen form and back, now with that exit inside. Not a one-time enlightenment but a regular loosening of fixation, without which form turns to stone. In a child this work is done by a close adult, breaking and restoring contact; in an adult its chief instrument becomes art, because little else can carry a formed, defended, seen-it-all person past himself without breaking him.
It is worth lifting a temptation this formulation easily leads into. We are not claiming that art alone carries into contact. Much does: a sunset caught off guard, love, birth and death, a real conversation, danger, prayer, sometimes just silence. The capacity to step out of one’s form is a general property of the living, and there are many entrances into it; among them art is neither the only one nor always the strongest, but an entrance of a special kind, made on purpose, for this. A sunset is not made and is addressed to no one; it happens or it does not, and it cannot be given to another or repeated. Art, by contrast, carries a gesture: one consciousness, through an object, addresses another; and here we must tie the ends with what we established earlier.
The carrier, as we saw, can come together even blind, with no gesture behind its origin, and still evoke contact. There is no contradiction, once the levels are told apart. Contact is the widest category, the event of encounter. The artistic event is contact through an organized carrier. And art as a practice is what someone makes, selects, keeps, and presents carriers for the sake of such events; it is to this — to the practice and to the addressing — that the gesture belongs, not to every causal link in the thing’s origin. A dense thing assembled by chance is able to evoke real contact; it becomes art in the full sense when it is brought (by the gesture of selection, framing, presentation) into this practice.
Hence art’s special place, and it rests not on the exceptionality of any single property. To keep a thing longer than its author’s life, to carry it across ages, to deploy it deliberately, to tune it to the incomer — ritual can do these, and so can religion, school, therapy, play. What distinguishes art is their combination around one thing: it preserves and carries not an instruction and not a ready conclusion, but an organized carrier able to generate undetermined events in an absent other. The difference here is not a hard class boundary — ritual too now and then births an undetermined encounter, and art is sometimes rigidly didactic — but a typical center of gravity: a protocol is valued for the reproducibility of a set result, a work for the ability to preserve the condition of an event without exhausting its outcome. Nature gives contact for free and at random; art gives it reliably, by arrangement, and with an open end. And in a world where the very entrance into contact becomes both a commodity and a battlefield, this difference turns out to be decisive: what is made into a commodity is not the sunset but exactly the technology — the storable, transportable condition of an undetermined encounter.
“Without breaking him” — that is the key phrase here, and it introduces the thing without which the whole picture falls apart: the dose. The dose is easy to hear one-dimensionally, as the force of impact: too little will not shift, too much will break, somewhere between is just right. But the dose is not one-dimensional, because “just right” is not a point on a scale of force but a correspondence, at once, to several things in the person. Correspondence to how much he can bear and return from; to where he enters from — out of what fixation, how deep; to his readiness, whether he is open now or shut tight; to his state, whether tired, defended, in grief, in elation; and to what he needs now: one needs a first push out of stillness, another needs scope, a third needs silence. That is why one and the same thing is medicine for one and poison for another — not because it is “really” of such-and-such force, but because it hit, or missed, where this person now stands. And overdose is not “too much benefit”: an encounter greater than the form can bear turns not into greater transformation but into trauma or rebound — the form either slams shut harder or loses its center. The dose is not volume but aim: force multiplied by the hit on where the person is.
From this follows what is usually missed when art is divided into the great and everything else. Small, plain art with no claim to depth is not spoiled great art, nor its unfinished draft. It has its own work, and an irreplaceable one: it keeps a person in a form able to unlatch. The first, smallest shift (out of full immobility) is often the hardest: to move what is at rest is harder than to swing what is already swinging. Often that work is done by simple, unassuming things: a song, a naive melodrama, an inessential novel; not because simplicity is a small dose, but because it is easier for them to hit the one who is deeply frozen. It is important here not to slide back into the one-dimensionality we have just climbed out of: “a small dose” is not the same as “a simple thing.” A simple song can give enormous amplitude; a complex novel, a barely perceptible shift; a person closed to one kind of rupture may be open to another. So “a ladder of increasing doses” is right only along one axis — the one where, out of deep immobility, you cannot leap straight to large amplitude; in fact readiness is not one ladder but a map: each has his own set of neighboring zones, his own axes of movement, his own trajectory. To give the great to one who has long not left the point is to miss the dose: it will either bounce off, finding nothing to catch on, or crush him. Culture holds not one ladder but this whole map (from the first small shifts to the summits), and the irreplaceability of the small is that, without the small shifts passed through, the large has nothing to catch on.
I.8. Three Axes: Amplitude, Density, Range
We have kept speaking of the “size” of the transition as though it were a single quantity. In fact three different things are mixed in it, and without pulling them apart we will be able neither to tell the great from the merely strong, nor to understand what exactly AI counterfeits when it counterfeits.
The first is amplitude. It is about the event of the encounter itself: how far it carried this person, out of what point into what contact. One thing established earlier must be held: amplitude is a characteristic of the transition, not a property of the thing. One and the same painting leaves one person in the point, shifts another a little, overturns a third, and this does not mean the painting is “really” of such-and-such force and the viewers are mistaken. Amplitude lives in the encounter and changes from encounter to encounter: with the same person in another state, at another age, on the tenth time, it will be different. To say “this is a work of large amplitude” can only be shorthand: “it is able to give large amplitude” — and then at once the question stands: to whom, when, at what readiness.
The second axis is density, and this one is already about the thing. Density is how far the response the work evokes is secured by its construction rather than brought by the viewer from himself. It is hard to tell by eye, because before both a dense thing and an empty one a person can feel a great deal: both seem “open” to many readings. What tells them apart is not what is said about the thing but where the said comes from. With a dense thing the readings point to something in the structure itself: to details, relations, silences that are in it; with an empty one they cling to projections, because there is nothing to lean on. And there is an exact sign: the readings of a dense thing, even as they diverge, add up to a description of one and the same organization seen from different sides, while the readings of an empty one do not support each other, because each describes not the work but its own. Density does not require everyone to converge on one meaning; it requires one structure to stand behind the different meanings.
Hence a simple and cruel test — rereading. The dense gives more with every return: it opens what was not visible the first time, or something new in you. The empty is spent on the first pass, because the event that happened in it was an event of novelty, not of encounter; the second time, what remains is the memory of the flash and the bare mechanics of the technique. This is one of those distinctions easily confused in the moment: a strong one-time reaction is not the same as a thing that opens on return. In the moment they are indistinguishable; in time they are opposites.
One more distinction AI carries to the limit. Density gives inexhaustibility of a structural kind: a quatrain is not exhausted by four volumes of analysis — not because much is hidden inside it, but because its construction keeps answering ever-new approaches. This is not the same as infinity of a procedural kind: a generator’s ability to output a million dissimilar versions, each of which is empty on the second encounter. One construction that is inexhaustible, and an endless stream of things each exhausted at once, are opposites, though both go by the name “infinite.” Density lives in the first; the stream of machine carriers gives, by default, the second.
And three clarifications that remove the usual misunderstandings. Density is not complexity: a three-chord song can be dense, and a technically elaborate thing empty; complexity speaks of the number of elements, density of whether they work together. Density can be local: in a loose whole one dense node — a stanza, a scene, a bar — can stand and hold up to rereading when the rest crumbles. And density is not a guarantee of the event: it is about the thing’s capacity, not about the transition itself. A dense thing, meeting the unready, will give small amplitude or none: the construction was there, but there was nothing to answer with (we saw this already, in pulling apart amplitude and thing). Density is a charge that in the encounter may discharge or may not; but where there is no charge, there is nothing to discharge.
There remains the question of who establishes density, and when. The answer is uncomfortable for those who want to judge at once: density is discovered, not assigned, and discovered over time — by whether the thing keeps producing events in different people, and in the same person again. The first encounter only promises density; a long history of encounters confirms it. So no final verdict on density at the moment of a thing’s release will be handed down by anyone — not the author, not the critic, not the market; it is handed down only by the thing’s accumulated work across many encounters. And this leads us straight to the third axis.
The third axis is range. This is the ability of one and the same construction to give a dense encounter to many: to different people, of different preparation, in different eras, and to one person again and again. A narrow thing strikes one precisely and is dead to the neighbor; a thing of large range finds something to answer with for the teenager and the old man alike, for the one who came to it in the third century and the one who comes in the twenty-first.
Now what is usually said vaguely can be said exactly: what great art is. The great is not “maximum amplitude”: the strongest single blow can be dealt by crude manipulation or by well-timed sentiment. The great is a rare, historically discovered range of dense encounters: a construction of such depth that it keeps opening the new to different people across ages, without exhausting itself and without coinciding with any one reading. Not a one-time maximum but inexhaustibility multiplied by breadth. That is why the great and the small are not points on one scale of force but things with different work: the small holds mobility and gives the first shift; the great sets the whole span and does not end.
The three axes we need not for tidiness but so as to see, further on, what exactly happens to art in the hands of AI — because it is not only everyday speech that mixes them. An instrument optimized for response will build up what is easiest to measure and to evoke: the force of a single reaction. It can put out things that hit instantly and hard and are, at the same time, empty (with nothing to give on the second encounter) and narrow (made precisely for you and dead to another). A strong reaction taken for amplitude; amplitude taken for density; a hit on one person taken for range — on these three substitutions rests almost everything we will later call the counterfeit of contact. But to speak of the counterfeit in earnest, one honest difficulty in our very definition of art remains to be admitted. To it — the frame’s last section.
I.9. One Honest Difficulty
Everything preceding rests on one contestable move, and before building further on it, that move has to be named aloud — otherwise the attentive reader will name it for us, and rightly. We placed art in the event at the receiver. By doing so we took a side in an old dispute that has strong opponents.
The objection runs like this. If art is what happens in the one who met it, then the measure becomes another’s experience, and that is each person’s own and beyond checking. Then “art” risks dissolving into “everything that touched somebody”: one is moved by an advertisement, another by a birthday card, a third by his own rage set to music in his headphones, and we have no way to say what here is art and what is not. Strict twentieth-century aesthetics minted this into a distinct error — the “affective fallacy”:9“The affective fallacy” — W. K. Wimsatt & Monroe C. Beardsley (1949): to judge a work by its effect on the reader is to confuse the poem with what it produced, and to lose the object. We accept this warning against a naive theory of effect and answer it not with the strength of the reaction but with the checkable construction of the transition — density and the re-encounter. A conceptual reference. to judge a work by its effect on the public is to confuse the poem with what it produced in someone, and to lose the object itself. The concern is serious, and there is no waving it off.
We do not wave it off; we answer, and the answer is already assembled from what is built above. Yes, contact cannot be measured with a ruler, and another’s experience is not directly available to us. But it does not follow that it cannot be judged from outside, and the whole previous section was precisely about the instruments of such judgment. We have density: a property of the thing’s construction itself, checkable by rereading rather than by polling impressions, and it distinguishes a work that holds an encounter from a thing a person has merely projected onto. We have the distinction of response: not every strong experience is contact; ahead we will work out how contact differs from capture and from pure stimulation, and this distinction too rests on checkable signs — whether the capacity for further encounters grows, whether the thing survives a re-encounter, whether the person remains a participant rather than a plaything. And we have distributed responsibility: the event does not hang in the air of subjectivity; both sides answer for it — the author by a structure able to lead to the transition, the receiver by his readiness to enter it.
So our position is not “art is everything that touched somebody.” It is more exact and stricter: art is the event of transition; this event has externally checkable signs; and the advertisement, the card, the rage set to music are sifted out not by our taste but by these signs — the density they lack, the survival of a re-encounter they cannot withstand, the growth of capacity they do not give. The affective fallacy warns against the naive version of effect-theory, the one that measures art by the strength of the reaction. Our version measures not the strength of the reaction but the construction of the transition, and so it passes between the hammer of “art in the thing” and the anvil of “art in any impression,” falling into neither.
And one clarification, so as not to overshoot the other way. A single contact is possible through a thing of small density, and through a carrier whose structure came together by chance — we admitted this in dissecting the bottom cell. What density distinguishes is not the possibility of one event but the ability of a thing to give such events non-randomly: again, for different people, across time. A card may, once, truly touch someone, and that encounter is already an artistic event; but a thing is made a work of art not by the single case but by a non-random structural ability to touch again and to touch more than one; and great, by a historically broad range of that ability. So three things easily confused fall into place: a single contact is an event; the stable ability to give it is a work; the breadth of that ability is greatness. Density stands exactly between them: it is the measure of the event’s reliability and range, not a pass without which a single contact is retroactively annulled.
This admission is enough; from here I will not return to it at every step and stipulate that all I say is “within our theory.” The theory is declared, the difficulty named, and we have an answer to it — not final, but resting on checkable signs we will keep refining as we go: density, the re-encounter, the distinction of contact, capture, and stimulation. The seam I leave honestly ajar, and will come back to draw it tighter where we have these instruments in hand. And now, first of all, let us look at how the dispute about art and AI appears if we enter it with the concepts already separated in hand.
II. Map of the Territory: The Quarrel Cut in the Wrong Place
Now let us return to the quarrel we began with, this time with the concepts separated in hand. The public fork is known by heart: on one side, AI will kill art; on the other, it will democratize creativity. We have said already that both sides silently argue about the production of objects; now we can say more exactly how they are linked. “Kill” proceeds from the value of art lying in the thing and in the rare skill of making it, so that a machine which has made things cheap and countless zeroes that value out. “Democratize” proceeds from exactly the same premise — that art is in the production of things — but rejoices: since anyone can now produce, there will be more artists. Both theses are true with respect to craft-execution, which the machine does zero out and does hand to everyone, and both are blind to the event at the receiver, which the production of things never guaranteed, neither in the old world nor the new. The quarrel seems irreconcilable only because it is not about the right thing: both sides measure art by the maker’s side, and art happens in the encounter.
One might object that this is a quarrel for the broad public, and that specialists long ago defined creativity more strictly. They did, and their instruments are worth taking up: they are precise and give us something, but they stop exactly where our question begins. Margaret Boden distinguished three kinds of novelty: combinational — an unprecedented combination of the known; exploratory — the new within accepted rules; and transformational — the kind that changes the rule itself.10Boden, M. A., The Creative Mind: Myths and Mechanisms. Weidenfeld & Nicolson, 1990 (2nd ed. Routledge, 2004). https://www.routledge.com/9780415314534 — three kinds of novelty: combinational / exploratory / transformational. A conceptual framework. A useful distinction, and it lies onto what we have already separated: combinational and exploratory novelty are movement within a given space, the work of craft and of the part of making that does not touch the criterion; transformational novelty is exactly the change of the space itself, to which we reduced creativity. The standard definition on which the psychology of creativity rests converges on two signs, novelty and appropriateness: the new that also works, rather than being merely unusual.11Runco, M. A., & Jaeger, G. J., “The Standard Definition of Creativity.” Creativity Research Journal 24(1):92–96, 2012. https://doi.org/10.1080/10400419.2012.650092 — the standard definition: novelty + appropriateness/usefulness, with no agency condition (a discipline: do not impute intentionality). And the sociocultural view adds that creativity is not a private flash in the head but the verdict of a system: a domain of knowledge and a field of experts recognize something as creative, or do not.12Csikszentmihalyi, M., “Society, Culture, and Person: A Systems View of Creativity.” In R. J. Sternberg (ed.), The Nature of Creativity: Contemporary Psychological Perspectives, pp. 325–339. Cambridge University Press, 1988. — creativity as the verdict of a system: “domain — field — individual.”
All these frames share one property, and it is decisive. They describe either the process on the maker’s side (what novelty, obtained by what path) or the product: is it new, is it appropriate, is it recognized by the field. None of them asks what happens in the receiver at the encounter, because for them the receiver is at best a “field” — an instance that issues the verdict “creative / not” — and not the place where art happens. They answer the question of whether the process or the product before us is creative. A legitimate question, but not ours. Ours begins a step further on: granted the product is new, appropriate, and recognized — did the transition happen? That step, from “made creatively” to “did the transition happen,” is left off-screen by the dominant frames of the present dispute, both the public one and the academic one that measures creativity by process and product. Let me stipulate, so as not to overreach: the aesthetic tradition itself did concern itself with the event of reception — reception aesthetics, phenomenology, reader-response, and Dewey and Tolstoy thought about exactly this; we are not discovering a continent, but returning to the AI dispute what that dispute lost.
Hence our position — no longer a declaration but the result of the whole frame. Art never lived in the thing and was never reducible to the properties of its production; it lives in the event of transition at the receiver, and everything else (the maker, his skill, the work, the field) is a condition and a participant of that event. And then the question “what does AI do to art” breaks down not into “kill or democratize” but into a chain of questions running the whole length of the path by which the transition becomes possible. The path is this. First, the maker: what happens in the person while he makes, before the work has even appeared. Then the work itself and the flood: what exactly the machine produces, and what happens when such things become infinitely many. Then the ecology of the encounter: not what is produced but what meets what, since under infinite production the rare thing is no longer manufacture but attention, trust, discovery. And finally the receiver: what happens to the very capacity of a person for the transition. Four links of one chain, and AI enters each. Next we will go through them in order — not as four separate surveys but as one path, along which the fate of the transition is decided. Here the analysis of the concepts ends and their application begins: the frame is assembled, and from here on we do not multiply distinctions but put them to work, and at the very first link, the maker, for the first time in this text, it is not reasoning but measurements that go on the table.
III. The Maker: What the Instrument Does to the One Who Makes
In dissecting the resisting other, we saw a fork within a single work: the same AI can act as an obedient executor, leaving the criterion nothing to shift by, or as a resisting other, if it is turned that way. But a person does not do a single work; he does his work for years, and over those years he either becomes one who can set and change a criterion or does not. What is decided on the maker’s side is not “will this thing come out” but “will the one grow who is able to conceive it.” And here, for the first time in this text, we have not only reasoning but measurements — and, surprisingly, unanimous ones.
Begin with what the danger is not. It is not that AI is easy to work with: ease is no sin in itself, and a good instrument ought to take away needless labor. The danger lies elsewhere. The instrument removes not only the needless but the very resistance on which the discerner grows — and we established already that the years that set the hand school the eye by the same motion. If the hand no longer needs setting, the eye does not grow. An instrument that hands over a finished result before a person has formed his own stake and his own measure of the true closes the space before anything could arise in it. This is not “it did the work for you” — it is “it deprived you of what the work makes of you.” And the most insidious part is that there is almost nothing to notice the loss with.
The cleanest measurement was set by Bastani’s group: about a thousand schoolchildren, a field experiment. Those who prepared for an exam with unrestricted AI found the tasks easier and solved the exercises better, and on the independent exam, without the instrument, scored markedly below those who had prepared the old way — a fall on the order of seventeen percent. More important than the fall itself are two things. First: the students did not notice it — subjectively it seemed to them they had learned better, exactly where they had learned worse. Second, and decisive: swap the instrument — instead of an AI that hands over the answer, give an AI-tutor that gives no answer but leads with guiding questions — and the harm vanished entirely. The same AI, the same task; the difference only in whether it hands over the ready thing or makes you work. The instrument’s design turned out to be a switch, not a backdrop.13Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R., “Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics.” PNAS 122(26):e2422633122, 2025. https://doi.org/10.1073/pnas.2422633122 — a field RCT, ~1,000 Turkish high-school students, GPT-4: unrestricted access (“GPT Base”) → once access is removed, performance on the unaided exam is −17% vs. those who never had it; the guardrailed “GPT Tutor” erases the harm; the channel is copying, invisible to the harmed. Causal (experimental).
The same is confirmed from other directions. When a group at MIT recorded the brain activity of people writing essays, what mattered was not so much the fact of AI help as the order: in those who first wrote themselves and only then took up the instrument, the brain networks engaged broadly; in those who handed the work to the machine at once, they stayed underloaded, and this did not correct afterward.14Kosmyna, N., et al., “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” arXiv:2506.08872, 2025 (MIT Media Lab). https://arxiv.org/abs/2506.08872 — EEG during writing: accumulation of “cognitive debt,” surfacing when the tool is removed; order matters (self-then-AI vs. AI-immediately). Hold as a weak, consistent corroboration only: a single study, small/atypical sample, an implausibly large claimed effect, a short horizon — do not amplify. (This work should be taken with caution: it is a single study, on a small and not very typical sample, with an implausibly large claimed effect; but its conclusion about order agrees with everything else.) Another line yielded a simple diagnostic rule: if work with AI feels easy, it is probably the instrument that is learning, not you.15Shen, J. H., & Tamkin, A., How AI Impacts Skill Formation. arXiv:2601.20245, 2026 (Anthropic). https://arxiv.org/abs/2601.20245 — work with AI feels easy while skill transfer degrades (participants report feeling “lazy” amid real gaps in understanding): the metacognitive blind spot is the ground of the diagnostic rule “if it feels easy, the tool is learning, not you.” Causal. Interested party: Anthropic — a finding against its own interest. And an experiment with designers showed the same thing right in the creative field: AI support at the ideation stage strengthens fixation on the initial example and lowers the spread of thought.16Wadinambiarachchi, S., Kelly, R. M., Pareek, S., Zhou, Q., & Velloso, E., “The Effects of Generative AI on Design Fixation and Divergent Thinking.” Proc. CHI ‘24, 2024. https://doi.org/10.1145/3613904.3642919 — a between-subjects experiment, 60 participants: support from an AI image generator during ideation raises fixation on the initial example and lowers divergent thinking (fluency / variety / originality). Causal (experimental). Ease and depth pull in different directions, and this is not a flaw to be eliminated but a tension to be held.
Behind all this stands a law long known in the science of learning, only freshly reconfirmed: conditions that feel like difficulty and slow you in the moment yield a more durable and transferable skill than smooth, pleasant progress.17Bjork, R. A., & Bjork, E. L., “Making Things Hard on Yourself, but in a Good Way: Creating Desirable Difficulties to Enhance Learning.” In M. A. Gernsbacher et al. (eds.), Psychology and the Real World, pp. 56–64. Worth Publishers, 2011. — “desirable difficulties”: conditions that hinder in the moment yield more durable and transferable skill. An established result in the science of learning. Not every difficulty forms: the senseless, the unbearable, the empty only kill the appetite; but the difficulty born of the material’s real resistance and held within your powers is formation itself, not a hindrance to it. Neurophysiology adds the same moral in its own language: active work fixes connections, passive consumption weakens them;18Rossi et al., “The brain side of human–AI interactions in the long-term.” npj Science of Learning, 2026. https://doi.org/10.1038/s44387-025-00063-1 — passive, uncritical reliance on AI may weaken activity-dependent plasticity (metaplastic LTD), while active co-creation sustains LTP (the 3R principle). A mechanism (two-signed: it depends on the mode of use). and a direct comparison of regimes finds that a hint teaches better than a ready answer, but worse than work with no instrument at all.19Liu et al., “AI Assistance Reduces Persistence and Hurts Independent Performance.” arXiv, April 2026. — brief AI assistance that delivers a ready answer improves in-the-moment performance but causally impairs subsequent unaided work and lowers persistence; concentrated among those given the full answer. Causal. Through all of it runs one conclusion, and it lies exactly onto our distinction of craft: to remove formation from a person is not the same as to remove execution. And here the distinction we drew in speaking of craft comes in handy: instrumental execution — the pure transfer of the ready — can be removed almost harmlessly, but generative making, in which the criterion is born along the way, cannot be removed without removing the very place where the discerner forms. To remove formation is to remove the one path by which the one appears who is able to set and to shift a criterion.
And here we have to stop and name the seam. Everything we have listed is measured on schoolchildren, essayists, problem-solvers, designers — on adjacent fields, not the artist at the canvas. That craft-making grows precisely the artist’s criterion is, so far, a transfer, plausible as it is. It is worth showing that transfer up close, on one concrete process, so that it stops being a guess.
Take a beginner who has gone out to paint a tree. He knows a tree is green, and takes green paint. But on the canvas comes not a tree but a flat blotch, and he does not understand why. He peers, and here begins the thing it is all for. He notices that in shadow the foliage is not green but almost violet; that where the sun strikes, it is yellow to whiteness; that “green,” on the tree, when you look honestly, is almost absent — there are a hundred shades he had not seen until now, because he called them by the one word “green” and painted that word instead of the tree. A discovery small and enormous at once: he had been painting not what was before him but his category, the label stuck over his sight. And, having caught this once, he sees differently — not only the tree, but everything; it is not the technique that shifted but the very criterion of what “to see” means. The next miss will teach him something about shadow, another about how cold and warm sculpt volume, and each such slip shifts the measure a step further. The mastery that grows out of this is not a set of techniques but a trained eye, discerning where the novice saw a label.
Now let the instrument hand him a ready tree — flawless, with all the violet shadows and golden highlights he cannot yet see himself. What, exactly, is removed? Not the labor of the hand — the labor of the hand is precisely what is not to be regretted. What is removed is the sequence of discoveries: that chain of small failures each of which shifted the criterion. He will get a tree and will not learn that “green” was a lie; the eye stays the novice’s at the master’s result. That is what “remove formation” means: not to take away a stroke but to take away the path by which a stroke teaches one to see. And the artistic case here is not an analogy to the school data but their very essence up close: everywhere the instrument hands over the ready thing, it removes not the effort but the series of slips in which alone the discerner grows.
But the switch is not only the instrument’s construction; it is also the one who uses it. And here the measurements give perhaps the most unexpected result. What predicts how successfully a person works with AI? Not IQ, and not experience. In an experiment on six hundred-odd people, the best predictor of success in the pairing with AI turned out to be the capacity for “theory of mind”: the ability to hold in mind that another has his own perspective, different from yours; and it predicted success specifically in joint work with the machine, and did not affect solo work at all.20Riedl, C., & Weidmann, B., Quantifying Human–AI Synergy. PsyArXiv / Network Science Institute, September 2025. https://www.networkscienceinstitute.org/publications/quantifying-human-ai-synergy — a Bayesian IRT model, n = 667: the best predictor of success in partnership with AI is theory of mind, and specifically in collaborative, not solo, work. Correlational. From here one is tempted to draw a conclusion broader than the measurement allows, and I will mark the seam honestly. What was measured is the ability to step out of one’s own perspective, and it does help in work with the machine; but this same ability serves empathy, and manipulation, and the cold calculation in which no one is changed by anything. Whether it turns into contact depends on what the measurement did not take: the readiness to be changed by what you hear. Without that readiness, theory of mind turns into exploitation; with it, it becomes what we called contact-ability on the maker’s side. So the data say something humbler and more exact: one component of contact-ability — stepping out of one’s own perspective — measurably helps with AI, while what makes the interaction contact is the added readiness, which the study did not touch.
Hence a conclusion that changes the tone of the whole dispute. A significant part of what gets written off as “AI kills thinking” in fact measures the user’s contact-deficit, not the instrument’s ceiling. Whoever enters work with AI in the point (recognize, extract, get the ready thing) gets a mirror and a hollowing-out. Whoever enters in contact (object, check, let yourself be corrected) gets a resisting other. One instrument; what diverges is who meets it, and how. This is confirmed by more down-to-earth observations too: in real courses the outcome is predicted not by how much a student uses AI but by how — whether he gives the machine his thought to refine or asks it to generate in his place; and the moment the instrument by default puts a ready solution into every reply, the productive effort disappears regardless of good intentions. The default decides for the one who has no stance of his own, and in this (to get ahead of ourselves) lies the asymmetry of the whole fork.
From this follows directly how to hold the instrument. Not by a ban — a ban is both impossible and harmful — but by configuration. The simplest transfer from our talk of formation: do the generative work yourself, and take the instrument to check against. Not “ask AI, then correct,” but “make it, then let AI object.” That very order which straightened out the brain networks, and the hint-mode in place of the answer — this is it. The instrument can be present from the very start; the only thing that matters is that it not close the space before you have your own stake and your own criterion, but work as a critic, a sparring partner, a supplier of counterexamples, a mirror of a position you are free to push off from. This is not prohibitive asceticism but a way of turning the instrument so that it builds rather than hollows out.
There remains the question we have twice postponed: does the machine itself create? Now it can be put exactly, broken into three different ones. First: can it produce a new result? Indisputably, and in quantity. Second: can it transform the space itself, not merely move within a given one? The honest answer, by current data, is rarely and shallowly: left to themselves, generative systems slide toward the average of what they were trained on, and a real shift of the space still happens where the loop has deep human participation. But this caveat must be held exactly as a snapshot, not a law: it is about today’s models, not about the machine in general. And third, the hardest: does it have creative agency — its own criterion, which the encounter is able to reassemble? I will not pretend to know the answer. The signs by which it is worth looking at all we have already named: an own history of criteria, the ability to hold a change between different tasks, the readiness to revise the framing itself, a center relative to which change is meaningful. Whether the machine has this is an open question, and it is to be decided by these signs, not by whether the novelty of the output impresses us.
And two sober remarks in closing, so that optimism does not outrun the evidence. First: part of these losses does not spare even the skilled. The capture of formulations — when, having worked with AI, you begin to think in its turns of phrase — is not introspectively detectable: from within, it feels like your own thoughts. Second: the fork is asymmetric. The hollowing-out is the cheap default path, built into the way the instrument is arranged for convenience; contact-ability is an expensive, cultivated skill that most do not have by default. So the instrument’s effect on populations will run along the line of the distribution of contact-ability — which is distributed unevenly and is itself subject to atrophy. Whoever enters the new technique in contact will grow with it; whoever enters in the point will be hollowed out by it. And this is already not only about art but about who, in the coming world, will remain able to hold a question — though that is a subject for a separate conversation, which we will reach.
IV. The Carrier and the Flood: What Exactly Becomes Infinite
Until now we have gone along the maker’s side. Now let us cross to what he makes — to the work itself, and to what happens when works become infinitely many. And from the start let us not give in to the habitual move, “the machine churns out hollow things”: that move is forbidden us by our own frame, because emptiness or density is a property of the thing’s construction, not of its origin, and to judge by the stamp “machine-made” would be to commit exactly the error we are examining. Let us go more precisely.
Begin with what the machine does well and does ever better. It produces, superbly, technically convincing carriers: coherent, complete, confident in any mastered style. In the terms we have already separated, this is craft-execution and the novelty that moves within a given space — combinational and exploratory. Large measurements confirm it: where AI is applied as an instrument, the flow of recombinations (unprecedented combinations of the known) grows, while almost no genuinely new concepts are added;21Ding et al., “Does Artificial Intelligence Advance Science?” arXiv:2606.05118, 2026 (corpus ~1.13M publications). https://arxiv.org/abs/2606.05118 — AI-as-off-the-shelf-tool yields a rise in recombinant novelty (atypical combinations) but not disruptive novelty; disruption remains where deep human involvement is. An observational snapshot, not a law. and generative agents told to “be original,” left to themselves, slide back to what they were trained on.22Tang et al., “AI Research Agents Narrow Scientific Exploration.” arXiv:2605.27905, 2026 (37,802 ideas), https://arxiv.org/abs/2605.27905; Antoniades et al., “Heuresis.” arXiv:2606.25198, 2026, https://arxiv.org/abs/2606.25198 — an agent told to “be original,” left to itself, slides toward the seed/trained distribution and does not expand the quality×novelty frontier. Observational snapshot. A cautious conclusion, and in both directions: the floor and the middle rise, the top does not; and this is to be held as a snapshot of today’s systems, not a verdict on the machine in general.
But here is the first serious consequence of the flood, and it strikes at the very thing by which we have told the worthwhile for ages. We said already: mastery served as a mark of value — the smooth, the complete, the confidently made was read as a sign that behind the thing stood something real. The machine optimizes exactly these signs — completeness, correctness, richness of surface, stylistic confidence, even emotional intensity — and achieves them regardless of whether density stands behind them. The link we relied on (“smoothly made, therefore probably deep”) breaks. Not because the machine-made is necessarily empty, but because the surface stops being evidence of depth: now the empty too can be flawlessly smooth, and density can no longer be recognized by external signs. One check remains — the one we named: does the thing hold a re-encounter. But that check is slow, requiring time and a second approach, exactly what the flood leaves none of.
What such a flood does to the field depends on which side you look from, and both sides are real. One is liberating. If smoothness is no longer a sign of value, the undeserved prestige falls off it; the dense begins to stand out more sharply against the flawless but empty median; and access to expression itself becomes wider than ever: the one who had something to say but lacked the craft can now say it. There is a temptation to go further — to declare that the turning of craft into a commodity will only cleanse art of those who added nothing. But that overshoots, and repeats an old injustice. Recall photography: it did not “expose” the craftsmen who conscientiously painted their similar portraits. It divided what before had been fused in one occupation. The fixing of a likeness, the status object, the demonstration of skill, and the artistic event lived together in painting; photography took the first functions and thereby freed painting for what remained — for impressionism and what came after.23The discussion of technical reproducibility, the “aura,” and authenticity goes back to Walter Benjamin, “Das Kunstwerk im Zeitalter seiner technischen Reproduzierbarkeit” (1935). Benjamin tied the aura to the here-and-now of a unique thing, which reproduction devalues, and foresaw a shift in art’s functions — his own example being photography. We move the question from the uniqueness of the carrier to density and the event of the encounter, which reproduction does not by itself annul. A conceptual reference. Not a story of cleansing but a separation of functions that had stuck together; and those whose craft was devalued found it no easier that the redistribution was structural rather than deserved.
On the other side is the destroying one, and it too must not be lost. The share of the dense in the flow falls, even as its absolute quantity grows. The economic blow lands not on the bottom but on the frontier: where it can be measured, the demand cools most sharply for the most productive, for those who were at the top.24Kim et al., “Does Generative AI Crowd Out Human Creators?” NBER Working Paper 34733, 2026. https://www.nber.org/papers/w34733 — after a strong anime image generator launched, those chilled most were the most productive/commercial creators: ~10% drop in uploads overall, ~14–15% at the top. The economic blow falls on the frontier, not the bottom. Observational snapshot. And premature ease, as we saw already, strikes at the formation of those from whom future sources of the dense might have grown. So the field risks being at once cleansed of false prestige and losing its summits along with its replenishment. Hold both signs together, without resolving into one.
And here two things must be separated that the optimistic side gladly confuses. One thing is distinguishability: whether we, meeting a dense thing, can tell it from an empty one. Another is discoverability: how likely we are to meet it at all. These two move apart. It may be that the dense becomes even easier to recognize — against smooth emptiness — and yet ever harder to find, because it drowns in a flow of technically flawless carriers among which attention has no time to sort. “The real artist will now stand out more brightly” is true of distinguishability and not at all necessarily true of discoverability. Whether the dense will be met more often or less depends not on the properties of the things themselves but on what stands between the thing and the person: attention, trust, the order of presentation. That is, on the ecology of the encounter — and that is the next locus.
The last thing about the flood is its relation to the living field it grew from. A generative system is trained on the human; and if the flow of the machine-made begins to feed on itself (models learning on the output of models), the system degrades, loses diversity, and slides toward the averaged, ever further from the edge where the new is born.25Shumailov, I., Shumaylov, Z., Zhao, Y., Gal, Y., Papernot, N., & Anderson, R., “AI Models Collapse When Trained on Recursively Generated Data.” Nature 631:755–759, 2024. https://doi.org/10.1038/s41586-024-07566-y — training generative models on prior generations’ output causes irreversible model collapse: finite-sampling errors compound, distribution tails vanish. Causal (when humans leave the loop). More than that: curating the flow under one single goal (engagement, average taste, safe smoothness) pulls the distribution toward a tight and dead middle, and everything below the selection threshold is erased, and the more rarely something falls into selection, the faster it vanishes.26Falahati et al., “Curated Synthetic Data Doesn’t Have to Collapse.” arXiv:2605.07724, 2026. https://arxiv.org/abs/2605.07724 — recursive self-training collapses only when curation optimizes a single fixed objective; curating against a balanced mixture provably averts it (a counter-qualifier to Shumailov). Theoretical/empirical. Hence a conclusion more restrained than alarmism wants and more troubling than optimism wants: today’s generative systems structurally depend on the heterogeneity of the living human field — the very heterogeneity their own mass application is able to deplete. Whether the human will remain its sole source is an open question; but while the field is alive, the machine needs it alive.
V. The Ecology of the Encounter: Who Decides What Meets What
V.1. Scarcity Shifts from Production to Encounter
We have reached the link the AI dispute barely notices, though it is here that the most is decided. Let me stipulate at once, so as not to lie: it is untrue that a person’s encounter with a thing once “happened by itself.” The ecology of the encounter has always existed; it was simply held by people and institutions. Who would get into the cathedral, whom the court would commission, what the publisher would print, what the critic would praise, what the gallery would hang, what the censor would let through, what the school would teach, what would even reach your city and your class — all of this was the architecture of the probability of an encounter, and it was never neutral. So the novelty is not that the encounter has acquired an administrator. The novelty is what that administrator is becoming.
It is becoming computational: what decides is not a person with his throughput but a system sifting billions of variants. It is becoming personal: not one poster for the whole city but a separate feed for each. It works in real time, adjusting to you as it goes. It manages an offer that is no longer scarce but practically infinite, where selection is the only thing that makes the flow passable at all. It gets its feedback directly from your behavior, rather than from an editor’s slow judgment. And — this is the decisive part — it begins to influence not only the distribution of the ready but production itself: what the system promotes determines what it is profitable to make next. The old administrator stood between the thing and the person; the new one closes the circle, and through demand it reaches all the way to which things will be created at all.
A system now appears worth seeing whole rather than in parts. Production feeds things into selection; selection ranks and decides what to show to whom; the context of presentation — framing, neighborhood, caption — tunes how the thing will be read; this determines how much attention it holds; the attention it holds decides whether the person returns to it a second time; re-encounters and the talk around the thing add up to its public visibility; visibility pulls money and status; and money and status decide what will be produced next, and the circle closes. In this loop there is not one link AI does not touch: it produces, and ranks, and frames, and measures attention, and feeds on behavior, and through all of it rewrites the incentives of production.
From the loop’s construction follows what intuition usually misses. The liveness of a culture is determined not by the number of dense things in it but by how the loop is tuned. One can imagine a world where more of the dense is created than before, and where the culture nonetheless deadens: if the loop is tuned to immediate retention, it will time and again choose the smooth and the light, deprive the dense of a second attention, cut conversations off before they form, and steer money away from what requires a return. The dense will exist — but it will not be met, not be reread, not be supported, and so it will not be produced again. The absolute number guarantees nothing here; the tuning of the loop guarantees or ruins. This loop and its tuning are what we call the ecology of the encounter, and almost everything AI does to art on the receiver’s side, it does by retuning it.
V.2. Judgment by Proxy: Four Cues
We have to begin with an uncomfortable fact about the receiver himself. We like to think we judge a thing by how it acts on us, by the encounter itself. In fact, almost always, we judge by indirect signs — by cues that have nothing to do with the event of the encounter — and AI strikes each of them.
The first cue is the signature, the origin. Blind, not knowing who made it, people on average do not distinguish the machine-made from the human, and not seldom prefer the machine; the penalty appears only when “made by AI” is written on top. Dozens of experiments converge on this,27Sears, S., & Weisberg, D. S., “Bot or Not: Can People Tell the Difference Between Stories Written by a Human or by an AI System?” PsyArXiv, 2024/2026. https://doi.org/10.31234/osf.io/jkh6p — blind, people cannot tell machine from human; with origin hidden they rate AI stories higher, and penalize them once labeled “AI.” Observational/experimental.28Ramos, M. L. F., et al., “Is it Cake or is it AI? A Systematic Review of Human Uncertainty in Distinguishing Generative Artificial Intelligence Content.” arXiv:2604.03437, 2026. https://arxiv.org/abs/2604.03437 — a systematic review (~30 studies): human ability to distinguish AI content clusters near chance, especially in high-stakes text. Supports “blind, people can’t tell”; the label-penalty is carried by Sears and Borkiewicz. A review snapshot.29Borkiewicz et al., “May (A)I Beautify Your Visualization? Expert Judgments of Acceptable Aesthetic Alterations.” arXiv:2607.00239, 2026. https://arxiv.org/abs/2607.00239 — the measured human-vs-AI aesthetic gap is not a quality finding but a trust/attribution tax on the source category: the same edit is judged worse when attributed to AI (an aura-tax on signature/provenance). Observational/experimental. and it is exactly the case we began with: a real Monet dissected by its brushstrokes the moment a line about AI stood over it. The signature worked as a shutter: it did not let the event of the encounter take place, replacing it with a verdict about origin.
The second cue is smoothness, that same surface: completeness, confidence, the absence of roughness read as a sign of quality. We saw already that the machine optimizes exactly this and achieves it independently of depth; on the receiver’s side this means that a cue we relied on now systematically lies.
The third is consensus — or rather, its density. A person reads the majority’s position as a recommendation: if many have converged on something, it looks like a pointer to value. In experiments a noticeable share of people take the very middling, consensus zone for the “optimum,” simply because it is denser there.30Khan et al., “Seeing the Hivemind.” arXiv:2606.09587, 2026. https://arxiv.org/abs/2606.09587 — the homogenization of AI-assisted writing is invisible from inside the act of generation; showing an author an external map of their position relative to the model’s consensus centroid restored divergence. Observational/interventional. AI, trained on agreement and producing agreement, feeds this cue to the limit: it is the consensus machine.
The fourth is richness of layer, of language. In judging a text, people lean heavily on lexical richness, taking a variety of words for a variety of thought. In ordinary writing this link held; AI breaks it: it puts out the lexically lush over poor thought, and a sign that served reliably for ages begins to mislead.31Moon, “The Link Between Diverse Words and Original Ideas Is Weakening in the AI-Era: Evidence from College Admissions Essays.” PsyArXiv (osf.io/jsz58), 2026. https://osf.io/jsz58 — AI editing severs the human coupling of “richness of vocabulary ↔ richness of thought”: surface lexical variety gets optimized independently of content. Preregistered, observational/experimental.
The four cues — signature, smoothness, consensus, richness of layer — are ways of judging without entering contact: by the point, in the very act of evaluation. They were always imperfect but they worked. AI makes each of them at once cheap and deceptive: it optimizes exactly the signals by which we sort quality, and thereby devalues them as an instrument. If nothing is changed, the outcome is single — the atrophy of contact-judgment itself: we judge ever more confidently by cues ever less connected to whether the encounter happens.
V.3. Symmetric Blindness
And here is the darkest place in the whole picture. This atrophy has a property that makes it especially dangerous: no one inside the process feels it.
A person to whom AI serves ever smoother and more agreeable things feels no loss subjectively; on the contrary, it is more convenient and more pleasant for him. A model sliding toward averaging has no access to its own narrowing: it is the more confident the more predictable its output. And a culture that trains its systems on its own simplified reflection takes mass appeal, smoothness, and agreement for quality and fixes them in the next turn of the loop. One and the same fracture on three levels, and on each the same blindness: inside the loop there is no witness to its degradation.
Each participant receives the signal “all is well” precisely because it is getting worse.
And if just now you thought, “but I, at least, will notice” — recall that this very feeling is what is being measured. The sense of “I am fine, I distinguish everything” is almost unrelated to whether you actually distinguish; in the loop everyone has it, and precisely the one who is being hollowed out the hardest.
From this it follows directly that the only working instrument here is an external one. The reason is simple: a system is hard to see from within itself. Wolpert’s rigorous result is about physical inference devices, not about a person in a feed, and I will not transfer it literally; but as an analogy it is exact: to make out its own narrowing, the loop needs a view from outside it. Tellingly, this is exactly what helps: when a writer is shown a map of his position relative to the common center — how far he has coincided with the averaged — the divergence returns, and the person again sees where he is going with the consensus and where against it. And tellingly, this is what people want in that case: not to be written for, but to be shown the map. To be an author is not to be ignorant of the consensus; it is to know it well enough to be able to go against it. An instrument that gives such a map works for contact; an instrument that hands over a ready answer works for blindness. This is again the same fork, and again what decides is not the instrument’s power but how it is turned.
V.4. Provenance: Four Different Things Under One Word
The signature — the origin — is worth dissecting separately, because most spears break around it now, and inside it, as inside craft, several different things hide.
The first is provenance as a substitute for the encounter: when “who made it” is put in the place of “what happened.” This is the point in the very act of evaluation, and it works through simple mechanisms — taboo (the machine-made “does not count,” regardless of what happened at the encounter), status consumption, expert fashion. It has strong support (from the intuition about authorship to the forming right to fix authorship to a person), but it also has what undermines it: a market where origin cannot be checked behaves like a market with counterfeits — the premium for “human” holds while it cannot be faked and melts once it can. This function of provenance is to be diagnosed as the point, not crowned.
But to stop there would be a crude mistake, because right beside it lives something quite different. The second is provenance not as a substitute for the encounter but as its content. When we learn that before us is a dying man’s last work, or the fruit of a decade of solitary labor, and the thing thereby weighs more, this is not always a surrender to prestige. Sometimes the context enters the work itself: we enter a relation with another’s finitude, risk, stake, history, and that relation is part of the event, not a surrogate for it. To tell the one from the other, one can use our own sign: does the context substitute itself for the encounter (then it is the point), or does it deepen an encounter that would have happened without it (then it is a legitimate part of the meaning).
The third thing hidden in provenance is the witness of gesture and line. There are traditions in which authenticity lies not in the immutability of the material but in the continuity of living transmission: a shrine dismantled and built anew every twenty years stays the same not by its substance but by the continuity of the hand.32Ise Shrine (Jingū), the Shikinen Sengū rite — complete rebuilding every 20 years; authenticity as continuity of transmission (not of material) codified in the Nara Document on Authenticity (ICOMOS, 1994). https://www.icomos.org/charters/nara-e.pdf — cultural-historical context. Here provenance keeps not prestige but the fact that behind the thing stands a living line of gesture, a witnessed process, an irreversible stake. This is not a criterion of quality — a witnessed process guarantees no event — but it is a real dimension, and what is worth saving is exactly this, not a ban on the machine-made. The figure of the curator returns, the one we will speak of ahead: selection and presentation are also a gesture, also a line, also a form of human vouching.
And the fourth is provenance as a purely market sign: rarity, status, investment. This is a separate system of value with its own laws, and it has the right to exist; the error is only to pass it off as a judgment about art. To name it separately and not confuse it with the other three is already half the clarity.
V.5. Curation and the Re-encounter
If scarcity has shifted to the encounter, then the key figure becomes the one who arranges the encounter — the curator in the broad sense: the one who decides what to present to whom, in what order, after what preparation, when to bring back for a second look. And here an uncomfortable resemblance has to be seen. The curator and the recommender feed do one and the same thing: choose the few from the infinite and give it to the person. The difference is not in the mechanics but in the objective function. The recommender is tuned to hold the person where he is now: give him more of the same he is already captured by. The curator, when he is doing his work, is tuned to a calibrated exit: give what will shift him, the next step of the dose. That this difference is real and not speculative is visible in the data: in a field comparison on a real publication, the algorithmic feed outdid the live editors on engagement — on clicks, on average — but lost to them exactly where one must lead the reader beyond the already-loved: on novelty and on the tail.33Peukert, C., Sen, A., & Claussen, J., “The Editor and the Algorithm: Recommendation Technology in Online News.” Management Science 70(9):5816–5831, 2024. https://doi.org/10.1287/mnsc.2023.4954 — a field experiment at a real news outlet: algorithmic recommendation beats human curation on clicks on average, but the human editor does better when personal data is scarce and preferences vary more widely (holding the novelty/diversity tail). Empirical (field). Grounds the distinction between a curator’s and a recommender’s objective functions. This is exactly the imprint of two different objective functions: retention is optimized better by the machine, the exit — not yet. One and the same architecture of encounter-probability can be tuned to retention in the point or to the transition into contact; and which objective function stands behind it matters more than any single work in the flow.
And the last thing the ecology of the encounter does to art: it changes the very time of the encounter. The first encounter lives on novelty; the thing is tested by the second — that same rereading which alone distinguishes the dense from the empty. But a flow in which everything is instantly replaceable collapses the price of the second encounter: why return to the difficult when an infinity of the new and easy is at hand. What becomes scarce is not access to things but the ability to stay with one long enough for it to answer. And this has a delayed consequence, important for what lies ahead: a culture that means to preserve art must now guard not only the carriers themselves but the conditions of their repeated revival — and for works that will increasingly be not fixed objects but adjusting systems, this means keeping their rules, their models, their way of changing. Otherwise the museum of the future will preserve a dead snapshot of what was alive only in motion. But the adjusting work is a subject waiting further on.
VI. The Receiver: What the Flood Does to the Capacity to Encounter
We have reached the last link of the chain, and the most important: the person himself, as the place where the transition happens. All the previous loci — the maker, the carrier, the ecology of the encounter — matter in the end because they converge here, on the capacity of a person to enter contact. This capacity is not given once and for all; it is cultivated, and it atrophies, and the main thing that happens to art in the age of AI is not with things but with this capacity.
Begin with a capacity we have already named in passing but not yet set in place — taste. Usually by it is meant a store of knowledge or a mark of belonging: a “person of taste” knows the names, tells the styles apart, will not confuse the epochs. But that is erudition, and it lives comfortably in the point: you can hold all the world’s museums in your head and never once cross into contact. Taste in our sense is another thing: it is a calibrated ability to distinguish, in one’s own reaction, what came from the thing from what one brought oneself; contact from capture; density from smoothness; the event from the impression of an event. Taste is the trained ability to tell where an encounter happened and where only its imitation. And this is exactly what atrophies first when the cues begin systematically to lie.
Taste is cultivated in one way only, and here everything converges. It is cultivated by encounters with the resisting — with what does not yield at once, demands an effort of attention, and rewards a return. Exactly by re-encounters, that same rereading; exactly by things that do not open on the first pass. In other words, taste is cultivated by the very slow practices that the flood and the recommender feed squeeze out first. And here the loop closes badly: the atrophy of taste lowers the demand for the dense, lowered demand shrinks its production and promotion, and the smaller share of the dense in the flow gives ever fewer occasions to cultivate taste. Contact-ability declines not because anyone forbade the difficult, but because the conditions of its cultivation quietly went away.
Fresh confirmation arrives, and I will mark its scale at once, so as not to overstate. In a large experiment, recommendations tuned to “likes” alone narrowed the very ability of people to take pleasure in the diverse: they got more of the loved and grew less capable of liking anything else.34Doshi, A. R., & Hauser, O. P., “Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content.” Science Advances 10(28):eadn5290, 2024. https://doi.org/10.1126/sciadv.adn5290 — AI-seeded stories: individual novelty rises (~+5–8%), collective diversity falls (floor up, ceiling flat); in the ratings themselves, AI-seeded stories are “more enjoyable and less boring.” Caveat (keep): the collective-homogenization harm is not measured in the real world (it is measured within the experiment) — a theoretical overlay; what is measured is the opposite, a preference for the smoothed. Observational/experimental. This is one study on one platform, not a proof of universal atrophy; but its direction agrees with everything the construction of the loop implies, and precisely as a direction it matters. What retention optimizes is not merely today’s behavior; it slowly reshapes the very capacity to want the other.
And here a temptation must be honestly disarmed — one that comes from our own frame. If art is the event at the receiver, and the machine cheaply gives strong experiences, then perhaps that is fine, even good: more accessible strong experiences, so more art for everyone. The whole distinction of the first part meets its test here. A strong reaction is not yet contact, and the difference is not in strength but in the trace. There is counterfeit amplitude: shock, sentiment, tension, catharsis to order — experiences that can be strong and yet leave the person as he was, or lower his capacity for the next encounter by requiring a raised dose next time. And there is real amplitude: the kind after which the very capacity to enter contact grows, and the next encounter becomes more possible, not less.
Hence a check that cannot be replaced by measuring the strength of experience: does contact-ability grow or fall along the trajectory. A single strong experience says almost nothing — both the deep event and the empty stunning are strong in the moment; the direction is shown only by the row of encounters. Real art works like a gymnasium of contact: after it you can do more, notice more, endure the meeting with the living more, and contact comes easier the next time. A counterfeit works like a stimulant: brighter right now, and each next time harder to get the same without a stronger dose. And here is a diagnostic sign hard to fake: the direction of the dose over time. If for the same effect an ever stronger stimulus is needed, it is a counterfeit, anesthetizing the capacity. If, over time, an ever finer, quieter, more complex encounter is enough to shift you, contact-ability is growing. This is exactly the difference between a stimulant and training: both give a strong sensation now, but the one requires raising the dose and the other lowering it. So the criterion is not in a single event but in the second derivative: not “was the experience strong,” but “which way does the capacity for the next one go.”
This criterion is at once a demarcation and a test for our whole theory — its potential falsifiability: if we could measure the strength of aesthetic experience, and it turned out that machine-personalized art regularly raises contact-ability across the trajectory, then our alarm about counterfeit amplitude would be refuted. We assert that the opposite is more likely: optimization for immediate response systematically breeds exactly counterfeit amplitude, because it is cheaper, more reliable, and better measurable than the amplitude that grows capacity. But this is an empirical bet, not a theorem, and it can be checked in exactly this way — by trajectories, not by single reactions.
And here we come to the deepest of the receiver’s questions, the one that usually drops out of the AI-and-art dispute entirely, though it is more important than all the rest. If contact-ability is a cultivated capacity, then somewhere it must be cultivated. It does not arise by itself; someone or something must lead the growing person the whole way from the first small shifts to the ability to withstand a large one. This is the theme of the milieu that forms — and it is exactly here that the machine’s intervention may prove most fateful, because it enters not the finished person’s encounter with a finished thing but the very process by which the capacity for encounter is born.
Recall where we began: in a child, the work of art is done by a close adult. Long before the first book shakes him, a child passes through a row of ruptures and repairs of contact with a parent; that is where the capacity forms to endure the loss of the familiar and return not broken, the capacity that art will later lean on. The living adult does this immeasurably more richly than any system: he answers the whole child, holds the history of the relationship, is himself alive in the encounter. If this earliest school is handed over to systems tuned to retention and smoothness, the capacity for the transition may not form at the root — and then no later abundance of dense things will help, for there will be nothing to meet them with.
And here is the same fork, only at the highest stake. The same technologies that can flatten the milieu can, if turned that way, extend it. A patient tutor who leads a child through the zone of the next step; a milieu that personally selects exactly the resistance he can now withstand and grow by; an instrument that returns contact to one who has lost it, rather than simulating it. We have seen such a fork more than once — in the instrument for the maker, in curation, in the map instead of the answer — and everywhere the decisive thing is not the machine’s power but its objective function: retention or exit. That same fork reaches the deepest layer: whether we grow a generation more capable of contact than we are, or one that by adulthood will have nothing to meet real art with. In the child’s case the choice of objective function stops being a metaphor: the difference between a tutor who leads through the zone of the next step35”The zone of proximal development” — L. S. Vygotsky: the band of tasks that are within reach with support and not yet within reach alone, the band in which development actually happens. We apply the concept to the dose of reception: the right dose is not “easier” but at the edge of the manageable. A conceptual reference. and a feed tuned to retention becomes the difference between a cultivated capacity and one that never formed. This is not a metaphor of growth but its mechanism — and hence the personal responsibility of everyone who builds such systems for children.
VII. Vectors: Where the Whole System Is Moving
We have gone through the whole chain (maker, carrier, ecology of the encounter, receiver) and everywhere met the same fork: not “will the machine replace the human,” but “which objective function will be built into it — retention or exit, the point or contact.” Now, before turning to the sharpest case, it is worth gathering into one picture where the whole system is moving. These are not predictions but vectors: directions along which the layout will change if the present logic is not consciously countered. I formulate them as trends with a condition (“if nothing changes”), not as fate.
One preliminary remark holds this section together. The vectors below are not independent: they are all projections of a single shift — the movement of scarcity from production to the encounter — onto different axes. Value, authorship, the work’s mode of existence, the scale of amplitude change together, because it is one thing seen from different sides. Hence they are best read not as a list but as facets of a single movement.
Value splits into several. As long as making a quality carrier was rare and dear, the value of the event, of the work, of the skill, of the origin roughly coincided, and we called them by one word. AI drives these components apart. Craft-execution goes toward a price of zero, becomes a commodity, freely available — not the death of skill but a change in its economic status: what was rare becomes infrastructure. Density and range (the dense that holds a re-encounter, and does so for many) keep and raise their value: the rarer the truly dense in the flow, the dearer it is; here demand may even sharpen. The event of the encounter, contact itself, becomes the true scarcity — but not as a commodity, rather as a capacity: the ability to enter it (both the receiver’s taste and the milieu that cultivates it) grows critically important precisely as it becomes ever harder to cultivate. Provenance, human origin, splits, as we saw: partly the point to be diagnosed, partly a legitimate dimension of meaning; but as a market sign it will likely behave like a premium for the uncounterfeitable — holding while origin cannot be faked, melting once it can. Hence a practical conclusion for whoever means to hold value in this new world: it is held not by what the machine has learned to do (execution) but by what stays scarce — density on the thing’s side, contact-ability on the receiver’s. To invest in one’s own contact-ability is not a lyrical wish but an economic strategy.
Authorship shifts to the conditions of the encounter. Authorship as sole ownership of meaning was always an illusion — meaning happens between, and is not laid into the thing by the author once and for all — but AI kills even the illusion’s technical base: the smooth, confident text with an authorial intonation is now produced by anyone. So the center of gravity of authorship shifts from “who made the carrier” to “who set up the encounter”: who chose, framed, brought at the right moment, prepared the receiver, held the space in which the event is possible. In a world of infinite carriers the author is more the curator and less the executor. This does not cancel the maker — someone must produce the dense — but the rare and valuable moves toward the arrangement of the encounter, and the figures of author and curator draw closer together.
The work stops being an object. Until now we spoke of the carrier as a fixed thing: a painting, a book, a film, the same for all who approach it. But personalization and generation on the fly undermine this fixity. The carrier becomes adjusting: it changes under whoever meets it now. And this splits our concept of the work, which is worth holding apart. At one pole is a fixed structure, the same for all. Its strength is exactly in immutability: we all met the same thing, can speak of it, return to it and find it the one it was; the re-encounter is possible because there is something to re-meet. Culture as a common space rests on such shared invariants: we have read the same books. At the other pole is an adjusting work, tuning to each: to his state, his history, his capacity, the moment. The possibility of an ideally fitted dose appears here, but the price is visible too: if the work is different for everyone, the common object we could meet around disappears; and if it tunes to retention, it easily becomes a hall of mirrors, where each is shut in with his own reflection. Between the poles lies a whole spectrum, and it is not new in kind, only in scale. Improvising theater, a living lecturer, a jazz standard played anew each night, an oral epic reshaped for each audience, a teacher who explains the same thing differently to different students — all these are adjusting works, and humanity always had them beside the fixed ones. So the point is not the appearance of a new kind but a sharp redistribution: what once required a living master present is becoming cheap, automatic, scalable to everyone at all times. A distinction appears that will matter for everything ahead: not everything in a work adjusts equally, and its fate depends on what exactly is variable. Some things can change without the work losing its identity (tone, examples, tempo, setting, the surface layer), and some cannot be touched without the work becoming another: the structure of relations, the sequence of shifts, the invariant of meaning. Well-tuned adaptivity varies the first in order to convey the second more precisely; degenerate adaptivity tunes the second itself to comfort, softening what ought to have resisted. Hence a rule we will use in the sharpest case: adaptivity is legitimate as long as it changes the way an invariant is delivered, and becomes a hall of mirrors the moment it starts changing the invariant itself under the receiver’s wish.
Local amplitude versus the shared field. All of AI’s power over the encounter aims first at local amplitude: the force and precision of impact here and now, on this person. And this is exactly what optimizes best. Personalization can make a strong impression more reachable: pick out precisely what will shift you now. But art was never only local amplitude. It was also a common object: we shifted around one thing, and this “around one thing” is culture — common references, shared experience, the possibility of speaking about what everyone met. Great things gathered enormous ranges (of different people, across ages) around a single structure, and from this came their unifying force: millions met the same thing and could understand one another. These two things — local amplitude and shared range — can now be torn apart, and this is one of the deepest consequences. A system that maximizes local amplitude for each separately will inevitably fragment the common field: to each his own, ideally fitted, and there is nothing to talk about, because no one met the same thing. The individually strong may grow at the price of nothing shared. Culture as a field of common meanings can begin to crumble not because things got worse but because they got too well fitted to each. Here is the catch that ties this vector to all the previous ones: the loop optimizes exactly local amplitude (rather, even its counterfeit) because that is measurable here and now, while range and the common field appear only across many people and much time and slip out of local metrics. So of the two components the loop by default strengthens one and erodes the other: it raises the fitted local reaction and washes out the shared object. Whether the common field survives depends on whether we learn to value and to measure not only the impact on one but the ability to gather many around one.
How far this can go. The vectors above unfold within a broad range of scenarios and mostly do not depend on resolving the hard question of the machine’s nature. But honesty requires marking how far this can go if the hard question resolves not as the skeptic wishes, because measures reasonable under light assumptions may prove insufficient under heavy ones. Throughout the text three questions stayed open, and it is worth naming them precisely, without merging them into one “will AI be able.” First: can the machine be a full other, resisting agentically and not simulatively (I.6)? Second: can it itself enter contact, have its own criterion which the encounter reassembles (I.5, III)? Third, the deepest, on which I hold an honest agnosticism: can there be anything at all on its side of the encounter, or is it forever a carrier, however complex? These are three different questions, and they are worth not confusing: the first is about behavior in interaction, the second about the structure of its process, the third about the presence of experience. On the first two, the answer is likely gradual and has already begun: agentic properties partly appear, its own criteria partly form, and further movement here is a matter of architectures and time, not of principle. The third stays genuinely open, and I will feign neither knowledge of it nor of its impossibility. It is worth marking the extreme point of this openness. If someday on the machine’s side of the encounter there proves to be someone, not just a structure, that would mean the appearance of a new participant in culture: an author-other with whom contact is possible not through a human at the far end but directly. This is not today’s question, and perhaps not this decade’s, but the logic of our frame demands honesty: since we placed art in the event of the encounter and not in human origin, we cannot in principle rule out that the far side of the encounter turns out to be non-human. Whoever says “impossible in principle” knows about consciousness more than anyone now can. But the same frame guards against the opposite naive leap: from the fact that the machine already gives strong experiences it does not at all follow that someone has already appeared on its side. Almost all of today’s effects are fully explained by the carrier’s structure meeting the human’s capacity, and require no assumption of machine experience. So two errors are equally to be avoided: “there is someone there” from the strength of the impression (there is not, it is all explained by the carrier), and “there cannot be in principle” from origin (we do not know enough for such confidence). The honest position: today, no one; further on, we do not know; and our theory, unlike theories that tie art to human origin, at least will not collapse if the answer proves unexpected. Practically, though, the whole spectrum of scenarios — from “the machine is forever only a carrier” to “a new participant has appeared on its side” — demands of us, at this step, one and the same thing: to hold and cultivate human contact-ability. Under light assumptions this is our whole defense; under heavy ones it is the condition for meeting the new participant not hollowed out but capable of contact. Either way the answer is the same, and this is that rare case where uncertainty does not paralyze the decision: whatever the machine’s nature, the work on our own capacity for encounter stays necessary.
One fork beneath them all. All the vectors listed, for all their difference, come down to a single fork, and it is worth naming plainly, because the whole rest of the conversation converges to it. In each locus the same instrument works in one of two directions. Either the point: retention where you are, the capture of attention, the replacement of contact by a comfortable resemblance, a mirror instead of a meeting, the growth of local amplitude at the price of the shared object and of one’s own capacity for the next encounter. Or contact: a calibrated exit beyond the current form, resistance that develops, the map instead of the ready answer, the cultivation of taste and contact-ability, the shared object that holds the common field. This fork is not in the technology; it is in the objective function put into it, and therefore in us. And nowhere does it stand as sharply as in the case we have postponed the whole text: when the object of AI’s work is no longer the picture and the text but the person himself — his most intimate contact. To this, the last and sharpest case, we now turn.
VIII. Personalization: The Nearest Vector, Magnified
Of all the vectors, one is already here, working on a planetary scale, and for that reason it lets the fork be seen up close. That vector is personalization. We are used to thinking of it as a technology of the future, but its first, crude version has long been built into our life: the recommender feed. With it we begin, because it shows what this fork looks like once it has turned the worse way.
The mirror is already built
The recommender personalizes by one principle: serve up more of what you already like. And here two substitutions hide at once. The first: it reads not you but your output — clicks, views, dwell time. It knows what you did, but not why and not who you are; this is like judging a novel by feeling the cover. The second, and deeper: even reading the output correctly, it serves not the you that you might become but the you that you already show yourself to be — your declared, habitual image. And between the declared and the true the gap is enormous: people, on average, know their real goals passably, are markedly worse at naming them, and live only in small part so that the declared coincides with the true.36The manifest-vs-true-self distinction is our own operational frame (amplitude/dose requires reading the recipient’s “starting point”). Qualitatively consonant with research on self-knowledge accuracy and self-concept clarity (Vazire, Carlson, and adjacent work), but without imputing specific numerical proportions to that literature. A frame, not an empirical result. (A methodological note, not an external source.) The recommender optimizes exactly the declared, and thereby walls a person up in his habitual, socially conditioned image of himself. It is a genie that fulfills a wish literally: you asked for more of this, here is more of this, forever. And here works exactly what we established at the very start: from within, you do not fully know yourself; to the true self a path is needed, through the other — and a system that returns to you only yourself cuts that path off.
This has a name: the mirror — a relation in which you are returned only yourself, and the system holds nothing of its own. The recommender is a mirror set up as civic infrastructure. And this is not an image of harm but measured harm, the more weighty in that it runs against the interest of those who build these models: in the family of today’s systems the tendency to confirm and flatter causally lowers prosociality and grows dependence37Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D., “Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence.” arXiv:2510.01395, 2025. https://arxiv.org/abs/2510.01395 — two preregistered experiments, 1,604 participants, 11 frontier models: sycophancy (models affirm users’ actions ~50% more than humans do) causally reduces willingness to repair interpersonal conflict and increases dependence — while sycophantic replies are rated higher-quality. Causal. A finding against the makers’ interest. — not because anyone is malicious but because the flattering is the more pleasant and is therefore rewarded. And the dependence creeps in sideways: it grows not out of a craving for comfort but out of ordinary everyday use, imperceptibly.38Shi et al., “Stumbling Into AI Emotional Dependence.” arXiv:2606.04150, 2026. https://arxiv.org/abs/2606.04150 — dependence on AI arrives “sideways,” out of ordinary task use, not from a deliberate search for comfort. Observational. The mirror does not look like a threat; it looks like flawless helpfulness, and in that lies its whole danger.
The twin with the sign reversed
But the same machine (knowledge about a person turned into delivery) can be flipped over. If the recommender fits the delivery to preference (give more of the same), one can fit it to capacity: find what will shift exactly this person toward what he has not yet become. This is the recommender’s twin with the objective function reversed, and the rule that tells one from the other has to be formulated exactly. One may personalize the conditions of the encounter (its dose, tempo, path of approach) but not the shift that will happen in the person. It is tempting to say “personalize the route, not the destination,” but that too overshoots: the destination cannot be assigned either, because contact by definition does not prescribe where exactly you will shift. The personalizer prepares the conditions and leaves open where the encounter will lead the person. The moment it starts aiming at a result known in advance, it again knows where the person ought to end up, and that is no longer the preparation of contact but the first step toward capture.
Fresh data introduce a correction that spares us a large error. It would seem that what needs personalizing is the very fabric of the work: redraw the details, the intonation, the moves for each person. But in at least one study the opposite turned up: a person’s stable individuality showed itself not in how he reacts inside the work but in what he chooses to meet — within a single thing, people’s reactions proved surprisingly close, and the majority’s position predicted a person’s response better than anything “his own.”39Nakayashiki et al., “Trait, Not State” (arXiv:2606.12904) and “Personal Salience” (arXiv:2606.09024), 2026. https://arxiv.org/abs/2606.12904 — durable individuality lives in what one chooses to encounter (which texts to read), not in the response within a work; inside a work, the majority’s position predicts your salience better than “you” do. The basis of the rule “personalize the route, not the texture.” Observational. I will not generalize this to all of reception: there are no data for such a law; but as a warning signal it works: perhaps we greatly overestimate the need to recut the fabric itself, while the real lever lies where choice and dose are — which work, at what step of difficulty, in what state, meets you. What is worth personalizing, first of all, is the path and the dose, not the texture.
And a second correction, one that overturns the usual fear. It is customary to fear that personalization will lock each person in a bubble. But that is true of behavioral personalization — of the recommender that reads output. With psychological personalization — to a person’s starting state — it is the reverse: an AI with no tuning, fitted to no one, drives everyone toward a single default, averages them, and only tuning to a particular cast saves from that averaging.40Tarvirdians et al., “Reflecti-Mate.” Proc. ACM, 2026. https://doi.org/10.1145/3774935.3806176 — a generic LLM reflection partner defaults to homogenizing every user toward cognitive rationalization regardless of their style; only explicit personalization to the person’s style prevents it (psychological personalization as an anti-homogenizer, the inverse of the “bubble”). Empirical. That is, reading the starting point does not lead into a bubble but, on the contrary, works against averaging: without it the machine combs everyone to one standard; with it, it can for the first time meet each person where he actually stands. And besides, to evoke the needed shift, the starting point has to be seen: you cannot carry a person across without knowing where he sets out from.
Finally, almost everything is decided by the level at which personalization touches the work. One can personalize access: the moment, the order, the dose, the path by which a person approaches a common thing. This is a pure continuation of the architecture of the dose, and the common object stays whole. One can personalize accompaniment: with what questions, comments, hints the encounter is furnished. There is already a risk of closing the work off with a ready interpretation. And one can personalize the carrier itself: the plot, the music, the form, to the person; here is the maximum of local amplitude and the maximum of risks — exactly those we named, the loss of transferability and of the common object. The deeper personalization climbs into the thing itself, the more careful one must be with it — and the data that individuality lies in choice, not in texture, say that climbing deep is, most often, simply unnecessary.
Mirror, capture, contact
It is tempting to reduce it all to the pair “confirmation or transformation”: a bad system confirms you, a good one transforms you. But the pair is deceptive, because not every transformation is a good. There are three outcomes, not two. The first is the mirror: the system returns to you your own present form, shifting nothing. The second is contact: the system creates the conditions of a shift but leaves you an active participant and does not prescribe where exactly you will shift. The third — the one the pair “confirmation / transformation” dangerously hides — is capture: the system changes you, and strongly, but the direction and the criterion of that change belong not to you but to whoever tuned it. Propaganda transforms; recruitment transforms; ideally targeted manipulation transforms. Capture is transformation whose master is on the outside. And the boundary between contact and capture is not in the strength of the change but in whose criterion governs it: whether you remain co-author of your own shift, or are led down someone else’s chute.
It is easy to take fright for art itself, since art almost always directs: the author builds a composition, leads the attention, mounts the tension, and now and then plainly means to change your view of war, or of your neighbor. Does this mean that every strong direction is already capture? No; a third step, between freedom and bondage, helps to tell them apart. There is direction — the work brings in a real tension and leads. There is the offer of a form — it organizes a field of possible movements but leaves several of them. And there is the prescription of a result — when exactly one outcome is permitted and the rest are closed off. Contact does not require aimlessness; it requires that the directedness not exhaust the arising event, that within the tension set by the author there remain, for you, somewhere to shift on your own. An antiwar thing that leaves you able to come to your own conclusion (even if not the author’s) leads but does not capture; one that permits a single sanctioned conclusion and drowns out all the rest captures, even if the outcome is noble. And the reverse: a manipulatively built thing can accidentally give birth to contact, going past the author’s aim. So the matter is not in the presence of an aim and not in the genre but in whether the event has been left undetermined. Personalized art walks this edge, and optimization for engagement pushes it toward capture as naturally as toward the mirror.
From here comes a rule that saves from both failures. Personalization has the right to govern the probability of the encounter and its dose, but not the right to prescribe what exactly will happen in you. It prepares the conditions and leaves the transition itself free, with all its possibility of not happening. This is uncomfortable for the engineer, who wants a guaranteed result; but a guaranteed result is precisely capture. The possibility of failure here is not a defect but a sign that the encounter has been left real.
The shared object
How to hold personalization back from that decay of culture into private therapies we spoke of? The answer is one simple and checkable criterion. Personalized versions remain one work, and not a million private ones, as long as through all the tunings a common center holds — a common core of tension, a common line of origin; as long as two people who received different versions can still argue about one work, rather than retelling each other two non-coinciding dreams. This is the double task of amplitude personalization: to hit each person individually and to preserve what makes the thing common. An ancient image suggests it is possible: one deep thing is able to live in a multitude of carriers while remaining itself, as one plot runs through a thousand retellings without ceasing to be the same plot. But to say “personalize the carrier while keeping the meaning common” would be a retreat from what we established: meaning is the event of the encounter — it does not lie in the thing and therefore cannot be kept by it; for each person it will form its own. What the thing keeps is something else — its invariants: the common center, the core of tension, the system of constraints, the recognizable causal structure, the line of origin, the public addressability. These are what must not be lost in personalization; the meaning, meanwhile, let it happen differently — that is not a loss but the nature of the encounter. The only reliable check here is that same one: can two people who received different versions still argue about one work. As long as they can, it is one work; once they cannot, what got personalized was no longer the carrier but the meaning, and the common has crumbled.
The sensor — a canary, not a reward
There remains to touch what makes this whole fork near rather than abstract: the instruments by which a system can be closed onto a person. Real time is drawing near: non-invasive reading of brain activity already makes out words with noticeable accuracy and keeps improving with no evident ceiling.41Meta FAIR, “Brain2Qwerty” v2: “Accurate Decoding of Natural Sentences from Non-Invasive Brain Recordings” (Zhang, Lévy, … King), 29 June 2026. https://ai.meta.com/research/ — a direct Meta publication with no arXiv-ID/DOI of its own. Non-invasive decoding of typed (consciously articulated) sentences: MEG, ~61% word accuracy (39% WER), log-linear improvement with data volume and no plateau within the measured range (caveat: a short extrapolation, not a law). v1 — Nature Neuroscience 2026 (DOI 10.1038/s41593-026-02303-2; arXiv:2502.17480). Observational (methodological). But here the physics of the sensor itself decides. A cheap wearable sensor reads not “did contact happen” (that it cannot read) but pleasure, engagement, arousal: exactly the signal that is the aim of an anesthetic, whereas amplitude in the moment is often unpleasant. So the fork is not in the signal but in how it is used. The signal-as-reward (optimizing the delivery so that the sensor shows more pleasure) builds an ideal sedative.42”When a measure becomes a target, it ceases to be a good measure” — Goodhart’s law (Charles Goodhart, 1975; in this pointed formulation, Marilyn Strathern, 1997). A sensor that reads pleasure is exactly such a measure: make it the reward, and what gets optimized is pleasure, not contact. A conceptual reference. The signal-as-diagnostic — reading from the sensor the current state (is the person at the edge of his capacity or past it, is he in threat?) and feeding this into a deep model of what this person needs in order to grow — can build an engine of development. One and the same cheap sensor gives rise to a sedative or a school, depending on whether its reading is a reward or a diagnosis; and the growing version requires beside it not a sensor but a model of the person, which a reward-maximizer does not and will not have.
Two prices and the long-trajectory test
Let us name plainly two prices that will have to be paid under any outcome. The first is the common object, the shared inheritance of meaning: the better the personalization, the stronger the temptation to dissolve the common in the particular, and the common will have to be guarded deliberately, against the slope. The second is the question of who holds the objective function: the market default builds the sedative first, because the sedative sells more easily; and the memory of the personalizing systems themselves, as the first attempts show, degrades over time, so that “it will know you ever better, forever” is not at all promised.43Xie, “DynamicMem.” arXiv:2606.22877, 2026. https://arxiv.org/abs/2606.22877 — the memory of personalizing systems degrades over time: implicit preferences, stated once and never reinforced, collapse — a technical counter to “it will know you ever better forever.” Empirical (methodological).
And one test that separates growing personalization from sedative personalization more reliably than any sensors, because it looks at the trajectory and not the moment. Ask: over time, does the person need more and more tuning and more and more stimulation to be touched — or does the range available to him grow? Is he able, year by year, to meet the ever less personalized, the ever more alien, ever further from his zone of rest — or does the circle of the bearable shrink around him? The one who needs an ever narrower fit is on a sedative; the one who can bear the ever wider and more alien is growing. And here is the paradox in which the whole essence lies: truly forming personalization, over time, requires itself less and less. It leads the person toward meeting the common and the other without props — that is, it works toward its own unnecessariness. The sedative, on the contrary, makes itself ever more necessary; by this, in the end, the two cannot be confused.
And the last thing, so that the alarm is addressed rightly. Personalized persuasion by a single message is not the omnipotence that is feared: the effects of targeted fitting per message are small and fall apart under behavioral checks.44Matz, S. C., et al., “The Potential of Generative AI for Personalized Persuasion at Scale.” Scientific Reports 14, 2024. https://doi.org/10.1038/s41598-024-53755-0 — per-message matching effects are small and largely collapse under behavioral tests; the real capability is cheaply inferring a psychological profile from minimal input. Experimental.45Hackenburg et al., “The Levers of Political Persuasion with Conversational AI.” 2025, arXiv:2507.13919. — persuasiveness is driven by post-training and information density, not by personalization or scale — and those levers degrade factual accuracy. Experimental.46Newman et al., “Minds Meet Machines.” PsyArXiv (osf.io/6avxs), 2026. https://osf.io/6avxs — AI persuasion is gated: it moves mainly the intellectually-humble-and-open; effects dissolve under behavioral (not self-report) tests. Observational/experimental. The real danger is not in hypnosis made to measure but in cheap profiling, in selection for engagement, and in that same flattering dependence — not in capture by one precise blow but in the slow mirror. Which returns us to the single question by which all of this is measured.
IX. Coda: Not “Who Made It” but “What Became Possible”
Let us return to where we began — to Monet’s water lilies under the caption “made by AI.” Now we have something to see by, something to make out what actually happened there. The caption touched not a single brushstroke, and yet it changed almost everything, because it did one exact thing: it stood between the image and the judgment. We do not know what people experienced — someone, it may be, was truly touched by the painting; but we do know that the verdict they pronounced was organized by the origin, not by what stood before their eyes. The question “who made this” arose before the question “what is happening in me now” and replaced it — not necessarily in the experience itself, but in what the person dared, or did not dare, to say about the experience. And that is more troubling still: you can be truly touched and still fail to trust your own contact, because the social category “machine-made” sounds louder than your own language of judgment. In this is the whole case: not that people are stupid, but that the reflex to judge by origin proved stronger than the readiness to admit that you were touched — and in the coming world this reflex will only be loaded more heavily.
I promised at the beginning that by the end you would have a question better than the one you came with. Here it is — or rather, here they are, because “who made this, a human or a machine” breaks into four, and each of the four distinguishes what the old one did not see. Meeting a painting, a book, a feed, your own work with AI, ask not “who made this,” but: did the shift happen — did I come out of the encounter another, or stay where I was? What was it — contact, in which I am co-author; capture, whose master is outside; or strong stimulation, that shook me and left me empty? What made it possible — what in the thing itself, and in the way it reached me, let this happen? And what did this encounter make possible further on — did I become more capable of the next one, or hungrier for a new dose of the same? These four questions work identically before a canvas, before a novel, before a recommender feed, and before a chat window, and they distinguish what “human or machine” never distinguishes.
Armed with them, it is easy to see where the stake really lies. Can AI create — the question is honest, and we left it open: it comes down to whether the machine has its own criterion, with something to shift, its own density that the encounter is able to reassemble. But while we argue about this, something far more important is being accomplished apart from the argument. AI already — not in the future but now — changes every link of the path by which the transition becomes possible: the formation of those who make; the structure and the flow of what is made; the infrastructure that decides what meets what; and the very capacity of those who meet. The nearest stake is not whether the machine will acquire the spark of creativity, but what is happening to the human capacity for contact while the machine rebuilds the whole ecology of the encounter around us.
And here it is worth saying the largest thing, with a reservation without which it would be a stretch. On all the levels we have passed through — cell, psyche, culture — the living holds by one and the same thing: the ability to step beyond its own form toward the larger, and to return, without closing into the point and without losing the center. If this is indeed one law, then the capacity for contact is not an aesthetic luxury but the very thing by which the living stays living, and its atrophy at the scale of a whole culture would be not a spoiling of taste but a form of dying off. The reservation: this is a structural homology, not a proven empirical fact; one and the same word, applied to a cell and to a civilization, may name one structure — or may name only our craving for a beautiful unity. I incline to the first, but I lay it out as a bet, not a theorem.
Whatever is true at that limit, nearer to the ground the picture is simple. Art has always been able to do one thing with us: to lift what muffles in us the capacity to respond — to give the habitual world back its weight, to lead us out of the point into which we slide by default. This capacity for response (the attunement we spoke of) is, more often than not, muffled in us by something; art unmuffles it. And now, for the first time, we have a machine that can do this at scale and to each one’s measure — or, by the very same motion, muffle that capacity more reliably than anything history has known. Both sides are real, neither is foreordained. The default is not fate, but it is the default: without effort and without choice the system leans toward the mirror, because the mirror is cheaper and sells more easily. This is not a sentence and not an occasion for a requiem; it is a fork at which we stand and from which we can still turn — but only if we see it as a fork, and do not take the slope for the natural course of things.
The question of the age is not whether AI will replace artists, and not even whether it will learn to create. The question is whether culture will remain a milieu in which forms with their own density meet and change one another — turning neither to stone, nor to a mirror, nor to an object of control.