The Frame Is the Work
AI art, disclosure, and the line that’s harder to draw than people pretend.
I was at Tin Town Cafe in Courtenay, standing in line for a coffee near the cards rack. The walls were hung with the kind of assortment a small-town cafe accumulates over a season — local photography, a few watercolours, mixed-media pieces with handwritten price tags. Most of what people were looking at were prints: cards, posters, reproductions.
A woman behind me in line was working through the rack. She picked a card up, held it at arm’s length, put it back. I asked her — knowing as I said it that it was an odd question to drop on a stranger — how she’d feel if she found out the cards she was looking at hadn’t been made by anyone. That they were AI-generated. That the artist whose name was on the back had never picked up a brush.
She didn’t take long to answer. It would disturb her, she said. She wasn’t a tech person; she didn’t know much about AI. But yes, that would change something. She’d be less interested.
I told her I’d feel the same way. Then I got my coffee and went outside.
The question that stayed with me wasn’t the one I’d asked.
How would she have found out?
She wouldn’t have known it, but she had walked into the middle of an argument the art world had been having for four years.
The first move — call it the spectrum — is the one the institutions have settled on. AI involvement in art isn’t a binary; it’s a continuum. On one end is a pure scan of a hand-painted canvas, faithfully reproduced and printed. On the other is a Midjourney prompt typed by someone who has never made anything before. Between them sits every step the digital era has already made: Photoshop, algorithmic art, generative code, GAN-based pattern work, the hybrid pieces shown at the Eastside Culture Crawl in which an artist generates a pattern with AI, projects it onto canvas, and paints over the projection by hand. Each step was once contested. Each is now mostly accepted. The line keeps moving. By this argument, the question isn’t whether AI is in the work — it’s how much, and to what effect. MoMA acquired Refik Anadol’s Unsupervised in late 2023; Tate Modern’s Electric Dreams exhibition (2024–25) framed AI as the latest chapter in a long history of technology and art-making; Christie’s “Augmented Intelligence” sale in March 2025 brought in over seven hundred thousand dollars and turned out a younger audience than the house had expected.
The second move is older and more philosophical. Conceptual art has spent half a century arguing that the instructions for a work are the work. Sol LeWitt’s wall drawings — executed by anyone who could follow his diagrams — are the canonical example. Boris Groys, in his 2023 essay “From Writing to Prompting” for e-flux, made the connection explicit: a prompt is a kind of instruction, and prompting “brings us back to the classical figure of the author” — the figure conceptual art was supposed to have dispensed with. By this argument, prompt-to-Midjourney work isn’t a betrayal of art’s tradition. It’s a continuation of one of its most defended lineages.
The third move is more practical, and is the one collectors have already absorbed. Most successful contemporary artists don’t physically make their work. Jeff Koons signs sculptures executed by his studio. Takashi Murakami signs paintings produced by Kaikai Kiki’s factory of assistants. Damien Hirst’s spot paintings are mostly painted by other people. Claire Bishop named the practice in a 2012 October essay — “delegated performance” — and the art world absorbed the implication: authorship is signed, not always touched. If a sculptor designs a piece, has it manufactured in China, and signs the result, that is their work. AI is, under this framing, a fabricator like any other.
These three arguments are the load-bearing walls of every defense of AI art now being made by major institutions. Each is intellectually serious. Each was settled long before AI image generation existed. The work that has been done has been done.
But the most interesting artists working with AI right now are doing something the three moves don’t quite cover — and what they’re doing makes the woman’s question sharper, not softer.
Consider an artist who does this. They train a model on their own corpus — every painting, every drawing, every sketch they have made over twenty years. They fine-tune it on their style, their palette, their recurring motifs. They generate a thousand permutations. They pick the dozen they think are best — the ones most recognizably theirs, the ones that move something. They print them, sign them, sell them.
Under each of the three moves from a moment ago, this is defensible.
The training data was theirs to begin with. The labor was delegated, in Bishop’s sense, but the authorship is signed; this is what Koons and Murakami already do, just with a different fabricator. The prompt is an instruction in Groys’s sense; the curation — the act of selecting the dozen out of the thousand — is exactly the kind of judgment the conceptual tradition has been calling “the work” for fifty years. Every philosophical resource the art world has spent half a century building is available to them. There is, by the existing standards, no clean argument against what they’ve done.
This isn’t hypothetical. It’s roughly what Holly Herndon and Mat Dryhurst have been doing for five years. Holly+ — an AI clone of Herndon’s voice released in 2021 — lets other artists use her voice as an instrument, with her consent. xhairymutantx, their Whitney Biennial piece in 2024, used a model fine-tuned on Herndon’s likeness. The Call, their Serpentine North show in 2024–25, was built on a custom dataset of fifteen UK choirs whose members co-owned the resulting model under what Herndon and Dryhurst call a “data trust.” They are not hiding what they did. They have published the protocol. The choirs are credited.
Refik Anadol works the same vein, if more spectacularly. Unsupervised, the piece MoMA acquired in 2023, was trained on 138,151 images from MoMA’s own collection. Sougwen Chung trains models on her own drawings and then performs alongside a robotic arm that draws in her style with her. Anna Ridler hand-photographed and hand-labeled ten thousand tulips for Myriad (Tulips). The pattern is the same across every AI artist the major institutions take seriously: a consented dataset, a legible protocol, a named human in the chain. The custom-trained model is the form that survives critical scrutiny.
So here is the test. If the version of this practice that MoMA puts on a wall is defensible, what about the version where the artist does exactly the same thing but doesn’t say so? The corpus is still theirs. The model is still theirs. The curation is still theirs. The signature is still theirs. The only thing missing is the wall label.
The work would be philosophically identical. The aesthetic object would be indistinguishable. Under each of the three moves, both versions clear the bar.
And yet they don’t feel like the same thing. The difference is not in the artifact. That difference is what this essay is trying to name.
The difference is disclosure.
The pattern shows up everywhere by now. A lawyer drafts a brief with Claude and doesn’t mention it. A student writes a college essay with ChatGPT and submits it under their name. An accountant runs a model over a client’s books and presents the output as analysis. A journalist’s draft, written by a tool, appears under a byline that doesn’t say so. In every case the artifact might be defensible if the practitioner stood by it openly. In every case the practitioner does not.
The work is hidden. And the reason it is hidden — across professions that have almost nothing else in common — is that disclosure changes the social object even when it doesn’t change the artifact. A brief that opens “Drafted with Claude, reviewed and edited by counsel” is a different professional act than the identical brief that doesn’t say so. The ink on the page is the same. The transaction is different.
The research on this is considerable, if mixed. Shown work without being told who made it, people often can’t reliably distinguish AI-generated images, poems, and music from human ones — and sometimes prefer the AI versions. Told who made what, their preferences shift. Chamberlain and colleagues at Goldsmiths found this pattern for visual art in 2018; Köbis and Mossink documented it for poetry in 2021, using GPT-2 verse against professional poets including Maya Angelou. The picture in music is more contested — Philippe Pasquier’s lab at SFU has been investigating it since 2016, and findings there have been inconsistent. But the asymmetry that keeps showing up is one of task type: the bias against AI is sharpest in domains where the output was supposed to come from someone — art, poetry, music, creative writing — and weakest where the output is judged functionally, like code or analysis. The disclosure changes the object only where the object was meant to be expressive. The work the disclosure does is happening in the viewer, not in the work.
This is the move the art world has been spending four years not quite making. The three arguments above — spectrum, conceptual lineage, delegated performance — are arguments about the artifact. They tell us why a piece of AI-involved work can count as art. They tell us nothing about whether the artist owes the buyer an account of what they did. The philosophical case for AI art is largely complete. The procedural case has barely begun.
And the art world’s version of the disclosure problem is sharper than other professions’ in three ways.
First, art value is almost entirely authenticity. A forgery and an original can be pigment-identical, and one of them is worth tens of millions of dollars and the other is worth what it costs to ship. Nothing in the artifact accounts for the gap. Provenance and attribution do all the work. A Picasso is not a better-painted painting. This is unlike law, where a brief either persuades a judge or it doesn’t; unlike accounting, where an analysis either matches reality or it doesn’t. In art, the disclosure question isn’t a side issue to the artifact’s value. It is the artifact’s value.
Second, the art world already has more disclosure infrastructure than most fields. The U.S. Copyright Office requires AI-generated content above a minimum threshold to be disclaimed in registration applications. The Colorado State Fair instituted AI disclosure rules in 2023 after Jason Allen’s Théâtre D’opéra Spatial won its emerging-artist prize the year before. The EU AI Act’s Article 50(2) labeling mandate enters into force in August 2026, requiring AI-generated output to be marked in both human-readable and machine-readable form. The infrastructure exists. The norms are being written.
Third, and here is where the asymmetry lives — none of that infrastructure reaches the wall of a cafe in Courtenay. It barely reaches the wall of a commercial gallery. It does not reach Etsy, or Society6, or the print rack at the back of a craft fair. The disclosure regime exists at the registration layer, at the institutional-prize layer, and at the platform-output layer. It does not exist at the layer where most people actually buy art.
The art world has spent four years arguing about whether AI can make art. It has not, to any meaningful degree, started arguing about how the buyer is supposed to know.
How do you find out?
There are three places you might look. None of them help.
The first is technical. The major AI image generators — Midjourney, Stable Diffusion, DALL-E, Imagen — embed varying degrees of metadata and, increasingly, invisible watermarks. Google’s SynthID encodes a signal in the pixels of its model outputs that survives many but not all transformations. C2PA — the Coalition for Content Provenance and Authenticity, backed by Adobe, Microsoft, the BBC, OpenAI, and a growing list of camera manufacturers — is building a cryptographic standard for tracking what was generated, where, and by whom. The technology is real. It is also, for the case at the cafe wall, useless. The metadata travels with the digital file. It does not survive being printed. The watermark is degraded by re-rendering, cropping, scaling, photographing, and printing — by every step that turns a generated image into a card you can pick up and put down. The forensic detectors (Hive, Optic, Reality Defender) are probabilistic classifiers; their error rate is higher than the error rate of a museum provenance researcher with a good eye, and they are getting worse, not better, as generation models close the gap. A print on a wall is a forensic dead end.
The second place you might look is institutional. The art world has developed, over centuries, the most sophisticated authenticity-verification infrastructure of any cultural domain. Catalogues raisonnés, exhibition histories, sales records, expert attribution, scientific analysis of pigment and canvas. The system works because the institutions have skin in the game: a wrongly attributed Rembrandt costs a museum its reputation and a buyer their money. But this infrastructure operates at the top of the market, where the stakes are millions of dollars per work. It does not operate at the cafe layer, where the stakes are forty dollars for a card and the only documentation is a name written in pencil on the back. Galleries do due diligence on emerging artists for forgery; they do not, in any organized way, do due diligence on AI involvement. Certificates of authenticity, where they exist, are issued by the artist — and an artist who would not disclose AI involvement on a wall label is not going to disclose it on a COA.
The third place is legal. The U.S. Copyright Office’s disclosure requirement applies to registration; the buyer of a print never sees a registration application. The EU AI Act’s Article 50(2) labeling mandate applies to the moment an AI system produces output; once the output is rendered to canvas or paper, the labeling obligation does not follow it. The Colorado State Fair’s disclosure rule applies to entries in the State Fair. None of these regimes reaches the point of sale for the overwhelming majority of art that changes hands.
So the verification gap is not a temporary problem that better technology will close. It is a structural feature of how art is made, sold, and consumed. The infrastructure exists at the layers where the artist interacts with institutions. It does not exist — and there is no obvious way to make it exist — at the layer where the buyer interacts with the work.
This is what the woman in the cafe was responding to. She did not say I object to AI art. She said it would disturb her. She did not need to know much about AI. What she knew was that she could not check, and that there was no third party she could appeal to if she wanted to. The flinch was not about the technology. It was about the asymmetry — that the buyer is being asked to trust without the means to verify, in a market where authenticity is the central determinant of value.
There is one more move to make, and it is the move the essay has been circling.
The standard frame on AI art treats the aesthetic question as separable from the disclosure question. Critics say the work is good or bad, beautiful or shallow, original or averaged, and they make their case on what is in front of them. Jerry Saltz looks at Anadol’s Unsupervised and calls it a half-million-dollar screensaver. Ben Davis looks at the same work and reads MoMA’s wall label as an insinuation that art history is “a bunch of random visual tics to be permuted.” Holly Herndon and Mat Dryhurst look at the same kind of practice and say the protocol is the work, the data trust is the work, the legible chain of human consent is the work. All three positions are honest. None of them is wrong on its own terms.
A fourth voice deserves mention. Hito Steyerl, writing from the European left, reframes the question as one of political economy — generative AI as “derivative” production built on “large-scale data theft.” Her critique is adjacent to this essay’s argument rather than continuous with it: she is asking about the system that produced the artifact; I am asking about the social act through which the artifact is presented. The essay’s claim does not require her claim to be wrong. It requires only that her question and mine be two separate questions, and they are.
But none of the first three survives the move the essay has just made.
If the difference between defensible and undefensible AI art is not in the artifact — if two pieces of pigment-identical work can be the same aesthetic object and yet be morally and procedurally distinct because one is disclosed and one is not — then the aesthetic question and the disclosure question are not separate questions. They are the same question, asked from two angles.
This is not a new argument. The art world has been making it, in other domains, for a century. A urinal in a bathroom is not Duchamp’s Fountain; the same urinal in a museum, signed and exhibited, is. A perfect forgery of a Rothko is not a Rothko, even if the pigment, the canvas, and the brushwork are identical, because what makes a Rothko a Rothko is not the pigment. The conceptual tradition that produced Sol LeWitt, the institutional theory of George Dickie, the cultural reading of Walter Benjamin — all of them rest on the recognition that the artifact is not the work. The social conditions of its production and presentation are part of the work. The frame is the work.
What is strange about the current debate over AI art is that the philosophical resources for this move are already in the art world’s hands, and the discourse has been slow to deploy them. Critics argue about the artifact as if disclosure were a secondary matter — a side question of professional ethics, separable from the aesthetic core. But the art world’s own history says otherwise. The aesthetic core includes the conditions of disclosure. An undisclosed AI print signed as if it were painted by hand is not aesthetically equivalent to the same print presented honestly. It fails differently. It fails as a different kind of object — closer to a forgery than to a contested artwork — because what it does is not just produce an image but enact a relation with the buyer, and the relation it enacts is one of asymmetry.
This is the position the essay has been building toward. The Saltz–Davis critique is correct that most current AI art is aesthetically thin. The Herndon–Dryhurst defense is correct that the strongest practitioners are doing something serious. Both can be true. But both are arguing about an artifact that is incomplete. The artifact does not include the social act of presentation. The social act of presentation is where the disclosure question lives. And the disclosure question turns out to be where most of the actual aesthetic disagreement lives too.
A disclosed AI work — Herndon and Dryhurst at the Whitney, Anadol at MoMA, even a small-time printmaker who writes “generated with Midjourney” on the back of a card — is a different kind of object than an undisclosed one. Worse and better are separate questions. Different is the point.
I should say something about my own practice.
This essay was drafted in conversation with an AI. The cafe scene, the question to the stranger, the line about disclosure changing the social object, the structural argument from the artifact to the relation — these are mine. The sentences are sometimes mine word for word and sometimes redrafted by a model and edited by me until the seam disappears. By the time it reads as a single voice, the line between what I wrote and what I edited has stopped being clearly visible even to me.
Under the argument the essay has just made, this matters. Not because it makes the essay less valid, but because it makes the essay’s own act of presentation a small example of what the essay is about. The reader, holding the finished text, has no reliable way to check. The disclosure that should change the social object — this was written with AI in the loop — would not have appeared without this paragraph. There is no infrastructure that would have surfaced it.
What I have done is not, by the essay’s own argument, wrong. It is unmarked, and unmarked is doing work I did not necessarily mean for it to do. Whether that work diminishes the essay or sharpens its claim by enacting it is a question only the disclosure can settle.
The woman in the cafe didn’t ask me what I thought about AI. She didn’t ask me whether the cards on the rack were generated or hand-drawn. She didn’t have a position on the spectrum, the conceptual lineage, or the delegated-performance tradition. She knew none of that and didn’t need to.
She said it would disturb her.
She was right to be disturbed. Not because the cards were AI-generated — I don’t know whether any of them were, and neither did she. She was right to be disturbed because the question is unanswerable from where she was standing. The artist whose name was on the back might have spent a week painting the original. They might have spent ninety seconds with Midjourney. They might have done something in between — generated a pattern, projected it, painted over it, signed it. There is no version of being a careful buyer at the cafe that gets her to an answer.
What she was responding to was not AI. It was a market that has stopped being able to tell her what she is holding. Her flinch was a small piece of a larger reckoning the art world has spent four years not quite having.
The question is not whether AI can make art. The question is whether the painting is allowed to say what it is.