What AI art exposes is how little we ever agreed about art

Benjamin was writing about circulation, not about souls#

Walter Benjamin's essay on the work of art in the age of its technical reproducibility, written in 1935 and published in French the following year, is the most abused text in this debate. It is invoked as a lament for authenticity; Benjamin in fact thought the decay of the aura a political opportunity, and the closing pages are about fascism and film, not the sadness of postcards. He analysed how a reproductive technology changes a work's social position: its distribution, its accessibility, the institutions that authenticate it.

Generative models do something he did not anticipate. Photography made works copyable; models make them re-derivable. An output is not a copy but a fresh sample from a distribution fitted to a corpus. The object of interest is no longer a thing but a chain — corpus, weights, prompt, seed, selection, edit — and every link in it is a place where responsibility can be located or evaded. Maedeh Norouzi and Jesse Prinz, in the same issue, argue that the right comparison is therefore photography rather than the brush: AI is a mediated latent medium, in which manipulation is indirect and the process largely imperceptible to the person doing it. Their list of its affordances beats the marketing — statistical idealisation, digital aberration, recombination, hypermnesia. The model remembers no particular thing and forgets nothing in general.

Barthes offers the wrong consolation, and elsewhere the right one#

"The Death of the Author", first published in English in 1967, is pressed into service here so often that it has become a licence: if meaning was never the author's property, the absence of an author is no loss. This misreads Barthes. His target was interpretive authority — the critic's habit of closing a text by appeal to biography — not who made the thing, or who was paid, or who consented. He removed the author from the reading. He did not remove the writer from the desk.

He is nonetheless indispensable here, by another route. In "Le grain de la voix", published in Musique en jeu in 1972, Barthes prefers the baritone Charles Panzéra to the technically superior Dietrich Fischer-Dieskau, on the grounds that in Panzéra one hears the body in the throat: the friction of the larynx, the specific tongue, the materiality of a person making sound. The grain is neither expression nor interpretation but the audible trace of a body doing something difficult.

That is the serious version of the thought that the imperfect human voice still matters, and better than the usual appeal to cracked notes as authenticity. A synthesised voice has no grain in Barthes's sense, not because it is too clean but because no body's limits are registered. Whether that absence is an aesthetic deficit cannot be settled from an armchair, and there is evidence.

Collingwood's test is the sharpest against machines, and proves too much#

R. G. Collingwood's The Principles of Art (1938) draws the line between art and craft that remains the most serious obstacle to machine authorship. Craft executes a plan: the maker knows the end before beginning. Art proper is the clarification of an emotion the artist does not yet understand. You do not begin with the feeling and find words for it; finding the words is how the feeling becomes determinate. A system sampling from a distribution has nothing to clarify, having no inchoate state to work out of.

The objection is strong, and it would be dishonest to pretend otherwise. It is also, notoriously, too strong. Collingwood's criterion disqualifies a great deal nobody proposes to demote: most commissioned portraiture, most workshop decoration, a fair proportion of Bach. A test that on a bad reading excludes the Brandenburg Concertos is telling us about Collingwood's Romanticism as much as about art. The residue survives and is narrower: there is a difference between making something because you have worked out what you meant and making it because it scored well, and that is not obviously a difference between humans and machines.

Dewey moves the question to the audience, where there is evidence#

John Dewey's Art as Experience (1934), out of the William James Lectures he gave at Harvard three years earlier, relocates art from the object to the transaction. What matters is an experience with a shape — a build, a consummation, a felt unity — and that happens in a person, not on a canvas. Dewey thus licenses a claim that irritates sceptics: if a listener is moved by a generated piece, the experience is not cancelled retrospectively by learning how the sound was made.

He also predicts the complication. Brian Porter and Edouard Machery, of Pittsburgh's department of history and philosophy of science, reported in Scientific Reports in November 2024 that 1,634 non-experts distinguished machine-written poems from work by Chaucer, Shakespeare, Dickinson, Eliot, Plath and five others at 46.6 per cent accuracy — below chance — and that in a second study 696 participants rated the machine poems more highly. Their explanation is deflationary and probably right: the generated verse is simpler, non-experts find simplicity legible, and they mistook difficulty in the human poems for machine incoherence.

Expertise changes this. Anna Yoo Jeong Ha and colleagues at the University of Chicago, at the 2024 ACM computer and communications security conference, found 180 crowdworkers did poorly while artists experienced at spotting generated images did well. Oliver Jacobs, Farid Pazhoohi, Grayson Mullen and Alan Kingstone, in Empirical Studies of the Arts, found something more awkward: with 250 participants rating works singly on a five-point scale the labels made no significant difference, but with 102 participants choosing between paired works, preference for the human-labelled item was strong. The method was generating the finding.

Labelling studies sharpen this. Wenyu Zhang and colleagues at East China Normal University, using EEG and near-infrared spectroscopy on 125 participants, found identical paintings rated lower on creativity, aesthetic value and preference when labelled machine-made, with P300 differences suggesting people attend less to what they believe a machine made. Yongquan Wang and co-authors, in Frontiers in Psychology in June 2026, report the bias roughly four times larger under controlled than automatic processing: a judgement, not a reflex.

What the studies jointly establish is narrower than either camp wants: audiences are poor at detection, they are not indifferent to attribution, and how much attribution matters depends on whether you ask them to compare. Dewey's transaction includes a belief about provenance among its ingredients.

Goodman makes the ontology question answerable rather than portentous#

Nelson Goodman's Languages of Art (1968), out of his John Locke Lectures at Oxford six years earlier, supplies the only clean tool here. An art is autographic if the distinction between original and forgery is meaningful — painting, where the history of production is constitutive. It is allographic if it is not, as with music, where a correct performance from a notation is the work and there is no forgery of a symphony, only a bad performance.

Generated images sit awkwardly in both categories, and the awkwardness is informative. There is no original hand, so the autographic test finds nothing to authenticate. Nor is there a notation: a prompt is not a score, since it does not determine the output, and the same prompt with a different seed yields a different picture. What exists instead is a record — weights, prompt, seed, sampler settings — neither hand nor score but a provenance log, and the legal fight below is an argument about whether such logs must exist and who may read them. Barcelona's congress on the human as image-maker circles the same ground.

Adorno's charge needs a better version than the usual one#

Theodor Adorno's account of the culture industry, worked out with Max Horkheimer in Dialektik der Aufklärung (1944) and sharpened in "On Popular Music" (1941), turns on pseudo-individuation: variety by standardised means, every song superficially distinct and structurally identical. The version usually aimed at generative music assumes its conclusion; the better version is measurable. Alexandre Declos, of the Université de Neuchâtel, argues in the special issue that text-to-image models exhibit systematic aesthetic bias — a standing preference for beauty, spectacle, saturation and symmetry — and that the trouble is not the preference but that it arrives as neutral output rather than a choice somebody made. That is Adorno's claim rendered as an empirical property of a system.

The counter-argument deserves equal room. Thomas Hofweber, of the University of North Carolina at Chapel Hill, argues in the same journal that the standard objections to machines producing work of extraordinary aesthetic value do not hold up. His mechanism turns on what he calls the gap: humans recognise aesthetic achievement well beyond what they can produce, and that asymmetry is what reinforcement learning from human feedback needs. He concedes the obstacles are formidable. It is the most serious technical case on the optimistic side, and one Adorno's framework cannot dismiss, since it does not require the system to understand anything.

Where the evidence tilts Adorno's way is institutional. Universal settled with Udio in late October 2025 and Warner followed with Udio and then Suno in November; Sony has settled with neither, and the American Federation of Musicians sued both labels on 5 June 2026, alleging its members' recordings were licensed without compensation or credit. A regime built on three catalogues is a plausible mechanism for standardisation, and it is being assembled without anyone deciding it should be.

Heidegger asks what a technology lets show up, which is the right question about style#

Martin Heidegger delivered "Die Frage nach der Technik" in Munich on 18 November 1953. His claim was not that technology is bad but that a technology is a way of letting things show up, and modern technology lets them show up as Bestand, standing reserve: the Rhine as an energy supplier, the forest as timber. His contrast is with technē in the older sense, a bringing-forth, which is why reading him as an opponent of machinery misses the point; our coverage of Gestell sets out the structure.

What generative systems let show up is style, as a retrievable and interpolable parameter. Before diffusion models, "in the manner of Hopper" was a description a critic might offer. It is now an address in a latent space, which a system can be told to visit and blend at a specified weight with somewhere else. That is a genuine disclosure, making visible a structure in the history of painting that was previously only inferable. It is also the conversion of a painter's life of looking into standing reserve. Both are accurate, which is the sort of thing Heidegger was good at and his imitators are not.

Boden's framework is the strongest case for machines, and its weak point is known#

Margaret Boden's The Creative Mind (1990) remains the best starting point for anyone who wants the pro-machine case stated rigorously rather than enthusiastically. Creativity, for Boden, is the ability to produce ideas or artefacts that are new, surprising and valuable. She distinguishes psychological creativity, new to the person having the idea, from historical creativity, new full stop, and insists the first is what explains minds. It comes in three kinds: combinational, connecting familiar ideas in unfamiliar ways; exploratory, finding unnoticed possibilities within an established conceptual space; and transformational, altering that space so previously impossible ideas become thinkable.

Its chief merit is that it makes the dispute tractable. Current models are plainly good at the first two, exploring their training distribution with a thoroughness no human could match. Transformational creativity is the test they have not passed, and for a structural reason — changing the rules requires representing rules as rules, and a system optimised to reproduce a distribution has every incentive to stay inside it. Boden herself does not claim computers are genuinely creative. She rehearses the objections and leaves the matter open, which is not where most who cite her leave it.

Training data is where the aesthetics and the ethics meet#

All of that is unresolved. The law is not waiting, and by September 2026 there is a record, messier than either lobby reports.

In the United States no court has yet decided whether training an image model on scraped artworks is fair use. Andersen v. Stability AI (N.D. Cal., 3:23-cv-00201) survived a second motion to dismiss on 12 August 2024, when Judge William Orrick let direct infringement proceed against all defendants, induced infringement against Stability and Runway, and Lanham Act trade dress claims, dismissing the DMCA claims with prejudice. Orrick accepted that the plaintiffs' works might be "contained in Stable Diffusion as algorithmic or mathematical representations". The case stayed in discovery through 2026, trial listed for 2027 and no class certified.

The clearest holdings concern text, and they point in different directions. On 11 February 2025 Judge Stephanos Bibas, a Third Circuit judge sitting by designation in Delaware, held in Thomson Reuters v. Ross Intelligence that Ross infringed copyright in 2,243 Westlaw headnotes and rejected fair use; the Third Circuit heard argument on 11 June 2026 and has not ruled. On 23 June 2025 Judge William Alsup held in Bartz v. Anthropic that training on lawfully purchased books was transformative fair use but that keeping pirated copies in a permanent library was not — the distinction that produced a 1.5 billion dollar settlement, approved in July 2026, covering 482,460 registered works at roughly 3,000 dollars each. Two days later, Judge Vince Chhabria granted summary judgment to Meta in Kadrey v. Meta on the record before him, while writing that "in many circumstances it will be illegal to copy copyright-protected works to train generative AI models without permission" and faulting the plaintiffs for not developing a market dilution theory. That is less a defence victory than instructions to the next plaintiffs.

In England, Getty Images v Stability AI [2025] EWHC 2863 (Ch) was decided by Mrs Justice Joanna Smith on 4 November 2025. Getty abandoned its training claim mid-trial, unable to show the acts occurred in the United Kingdom, and its output claim, Stability having blocked the relevant prompts. What remained was secondary infringement, and the court held that although an "article" under sections 22 and 23 of the 1988 Act can be intangible, model weights are not an infringing copy: they are "purely the product of the patterns and features" learnt during training. Getty won narrowly on trade marks, over iStock watermarks in earlier models, and has permission to appeal. Anyone reading the judgment as settling the training question has read the headline only.

The music publishers' litigation against Anthropic shows how little is fixed. A stipulation in early 2025 preserved guardrails against reproducing lyrics in outputs; Judge Eumi Lee denied a preliminary injunction on the training side on 25 March 2025 for want of irreparable harm; the lyrics claim survived dismissal that October; and a second action, filed on 28 January 2026, alleges torrenting of pirated songbooks and seeks 3.07 billion dollars. The patchwork character of AI jurisprudence is not a temporary condition.

Regulation is moving faster than adjudication. Under Article 53 of the EU AI Act, providers of general-purpose models must publish a summary of the content used for training, on a template the Commission finalised on 24 July 2025, and keep a copyright policy respecting rights reservations. The obligations applied from 2 August 2025; enforcement powers begin on 2 August 2026; models on the market before August 2025 have until 2 August 2027. A summary is not a manifest, and the guidelines excuse retraining where information is unavailable or the burden disproportionate.

On the output side the American position has been settled since the Copyright Office's second report of 29 January 2025: prompts alone do not confer authorship, because registration requires "sufficient human control over the expressive elements". Jason Allen's Théâtre D'opéra Spatial won the digital arts category at the Colorado State Fair on 29 August 2022 after at least 624 prompt revisions in Midjourney; the Review Board refused registration on 5 September 2023, finding the human contribution de minimis. Six hundred and twenty-four revisions is a good deal of deliberate work, and it bought him nothing.

Here the two halves meet. Aesthetics has spent two years asking what a work is and who its author is, settling nothing; the courts, obliged to decide, have answered the same questions under other names. Whether a model contains a copy is Goodman's autographic question. Whether a prompt constitutes authorship is Collingwood's question about plan and execution. Whether an output competes with its training corpus is Adorno's question about standardisation, asked by economists. And since audiences already treat provenance as part of the work, disclosure does more than supplement the aesthetics; it intervenes in them, because telling people what a thing was made from changes what they see.

Sofie Vlaad and Elliot Samuel Paul, closing the JAAC issue, name the two failure modes: dismissing this work wholesale is snobbery, praising it uncritically naïveté, and the virtue they propose, discerning openness, is the unglamorous middle. Laurie Anderson, awarded the Kyoto Prize this year, has spent five decades working with machines meant to replace the people using them, and the lesson of that career is that the interesting questions arrive once the technology stops being news. The question in the gallery will still be who made this. The harder one underneath, forced into the open, is why we ever thought that mattered.

Martin Heidegger, Michel Foucault