The AI art trial that did not happen, and the question it still poses

A story went round the legal internet this month that on 8 September a jury in San Francisco began hearing Andersen v. Stability AI, the artists' class action against the makers of Stable Diffusion, Midjourney, DeviantArt and Runway, and that twelve citizens would for the first time decide whether training an image model on scraped artworks is copyright infringement. A law-firm blog dated 6 September announced it. It did not happen. The docket of case 3:23-cv-00201, before Judge William Orrick in the Northern District of California, shows a lawsuit still deep in discovery, with no class certified, a sanctions hearing pending for October and a jury trial set for 1 June 2027. The 8 September date was real once; it was an earlier trial setting that the court moved, twice, before anyone wrote the blog.

The correction matters because the case is genuinely the one to watch, and the reason is philosophical before it is legal. Three and a half years in, Andersen has produced no ruling on fair use, or on whether a diffusion model's weights are a copy of anything. What it has produced is a record of artists trying to say, in the vocabulary of a statute written in 1976, what was taken from them. Set beside Thaler v. Perlmutter, which the Supreme Court declined to hear on 2 March 2026 and which fixes the rule that a machine cannot be an author, the two cases mark the edges of a law that has two words for creative property, copy and author, and is discovering that neither names the thing at issue.

What the docket actually shows#

The chronology is worth setting down, since the error came from a stale version of it. Sarah Andersen, Kelly McKernan and Karla Ortiz filed in January 2023, when generative image models were barely a year old. Orrick was unimpressed at first: on 30 October 2023 he dismissed most of the complaint with leave to amend, demanding specifics about how "compressed copies" of training images could be said to sit inside a model, and confining the copyright claims to the sixteen collections Andersen had actually registered. The amended complaint brought seven more named artists, Runway as a defendant and a good deal more technical detail, and on 12 August 2024 Orrick let the core of it through. Direct copyright infringement survived against all four defendants; induced infringement against Stability and Runway; and, against Midjourney alone, claims under the Lanham Act for false endorsement and trade dress. The claims under the Digital Millennium Copyright Act went out with prejudice, as did breach of contract against DeviantArt.

Then the case slowed. A scheduling order had trial for 8 September 2026. By the parties' joint case management statement of 13 January 2026 the date had already moved to 5 April 2027, and the plaintiffs were asking for four more months, citing among other things "difficulty of locating a qualified machine learning expert", which is the sort of detail one would not invent. On 2 February 2026 Orrick granted the request in part. The order, docket entry 409, closes fact discovery on 1 June 2026 and expert discovery on 18 October, makes class certification and summary judgment motions due on 11 November 2026 and hears them together on 17 February 2027, sets a pretrial conference for 26 April 2027, and a jury trial for 1 June 2027.

The late-summer docket reads like the middle of a fight, not the eve of one. Joint discovery letters of 24 and 25 August accuse Stability and Runway of violations and Stability of failing to preserve evidence. On 15 September DeviantArt opposed a plaintiffs' motion for sanctions, which Magistrate Judge Cisneros will hear on 20 October; on 22 September a further motion to modify the schedule went in. Nobody has picked a jury, and no class has been certified; until the November motion is granted, the ten named artists represent only themselves.

What the artists say was taken#

To get in the door under the Copyright Act one must allege a copy. The artists' works are among the roughly five billion image–text pairs in LAION-5B, the open dataset scraped from the web on which Stable Diffusion was trained; that much is not disputed. But the training images are not stored in the model. What is stored is a set of numerical weights, and whether those weights "contain" the images in any sense the statute recognises is the technical heart of the case.

Orrick's answer in August 2024 was that this is a question of evidence, not pleading, and the plaintiffs had pleaded enough, partly out of the defendants' own mouths. Emad Mostaque, then Stability's chief executive, had said of the model: "We took 100,000 gigabytes of images and compressed it to a two-gigabyte file that can recreate any of those [images] and iterations of those." A man who describes his product that way has some difficulty later telling a court it holds no copies. Orrick also reached for the statute's generosity about media, a work being "no less a motion picture" whether on videotape, videodisc "or any other tangible form", and held that if the plaintiffs' works may be contained in Stable Diffusion "as algorithmic or mathematical representations", that is "not an impediment to the claim". He added that whether the evidence supports the theories "will be addressed at summary judgment". The model-as-copy theory, in other words, is alive but has not been believed; it has been permitted.

Yet read the complaint, and the copy is not what the artists are grieving. What Karla Ortiz objects to is that a user can type "in the style of Karla Ortiz" and receive an image that trades on twenty years of her looking and drawing. And style is precisely what copyright does not protect. It protects expression fixed in a work; it leaves ideas, methods and manners free to all, which is why every art movement in history was legal. This is the gap the Lanham Act claim against Midjourney was drafted to fill. Midjourney's chief executive had circulated, on Discord, a list of some 4,700 artists' names, promoted as "various styles" the product could produce; the plaintiffs on it say this is false endorsement, an implication that they sponsor the tool. Orrick noted that courts have "almost uniformly said no" to making style a trade mark, warned against "extending trademark law into the domain of copyright law", and let the claim proceed to be tested. It matters structurally: fair use, the defence on which the text-model cases have turned, is a copyright doctrine, and it has no purchase on a claim about names.

The artists' injury, as they describe it, is neither the reproduction of any picture nor the loss of any sale. It is the conversion of a name into a setting, a person's manner into a parameter. Copyright's vocabulary can get at it only sideways, by insisting that the model is a warehouse of copies; trade mark law can get at it only by pretending that a style is a brand. Neither description is false, exactly. Neither is what happened.

Thaler, and the rule that a machine has no signature#

At the other end of the same year stands a case that is over. Stephen Thaler, a computer scientist, tried to register a picture called A Recent Entrance to Paradise, listing as its author his own system, the Creativity Machine, and claiming the copyright as the machine's owner. He has been consistent that he neither prompted the system nor altered what it made. The Copyright Office refused registration; the district court in Washington upheld the refusal; and on 18 March 2025 a panel of the D.C. Circuit, Judge Patricia Millett writing for herself, Judge Wilkins and Senior Judge Rogers, affirmed. The Creativity Machine "cannot be the recognized author of a copyrighted work because the Copyright Act of 1976 requires all eligible work to be authored in the first instance by a human being". The reasoning is textual and rather charming: "Machines do not have property, traditional human lifespans, family members, domiciles, nationalities, mentes reae, or signatures."

Thaler petitioned the Supreme Court on 9 October 2025, asking "whether works outputted by an artificial intelligence (AI) system without a direct, traditional authorial contribution by a natural person can be copyrighted". It was considered at the conference of 27 February 2026 and denied on 2 March. The D.C. Circuit's rule therefore stands, and it is a narrow one. Millett was explicit that the court was deciding only the case in front of it, in which the human had disclaimed all contribution, and that the requirement "requires only that the author of that work be a human being—the person who created, operated, or used artificial intelligence—and not the machine itself". How much a human must do to count is not answered by Thaler. It is addressed, for now, by the Copyright Office's second report on artificial intelligence, released on 29 January 2025, which holds that "prompts alone do not provide sufficient human control to make users of an AI system the authors of the output", while accepting that creative modification and selection can earn protection for what the human contributed.

The philosophical commitment here is older than the statute. The Copyright Office and the D.C. Circuit have adopted, without calling it that, an intentionalist theory of authorship: a work is the trace of a mind that meant something by it, and where no mind meant anything there is no work, only an object. The industry has not minded: an output no one can own is also an output no one can be sued over, and Thaler, who has pursued the same point through patent offices on three continents, has fought largely alone.

Two words, and the thing between them#

Put the cases side by side and an asymmetry appears that neither court has confronted. Thaler settles that what comes out of the machine, unassisted, has no author, and so belongs to no one. Andersen asks whether what went into the machine still belongs, in some legally cognisable sense, to the people who made it. The output is an orphan; the input, the artists say, has too many parents, five billion of them compressed into a file the size of a film. The model between them, the thing sold, is a copy to the plaintiffs and a statistical abstraction to the defendants, and an English court has already sided with the defendants under its own statute. We have surveyed the wider run of holdings, and the aesthetic theories they put under pressure, and will not repeat it here.

What the two American cases show, taken together, is that the law's pair of concepts was built for a world in which taking and making were distinct acts. Copying was the wrong, authorship the right, and the space between them was called influence, which was free. Every painter learnt by looking at other painters; the law never had to say what looking was, because it left nothing behind that could be counted. A diffusion model is a machine for looking that leaves something behind, and the litigation is a long argument about what to call the residue. Call it a copy and you must show, at summary judgment, that Ortiz's pictures can be got back out of the weights, which no one has yet demonstrated in a courtroom. Call it influence and you have said that the only difference between an apprentice and an industry is scale, which is true and not the whole truth. Call it, with the Lanham Act, a misuse of her name, and you have at least located the injury where the artists locate it, at the cost of admitting that copyright, the law of works, has nothing to say about it.

The intentionalist theory behind Thaler cuts here in a way its advocates might not welcome. If authorship is the mark of a mind that meant something, then the artists' claim is not really that their pictures were copied but that their minds were: what the model learnt from Ortiz was not this or that image but the pattern of her choosing, which is the nearest thing to her that survives outside her. There is a long tradition, from Kant's 1785 essay on the wrongfulness of unauthorised reprinting through the continental doctrine of moral right, which treats a work as an extension of a person rather than a parcel of property, and on that view the harm in Andersen is a harm to persons, of a kind the American statute, with its market-centred fair use test, was not designed to weigh. The original complaint did plead right of publicity claims under California law; Orrick dismissed them in October 2023 for want of any allegation that the defendants had used the artists' names to advertise, and the Lanham Act claim over Midjourney's list is what took their place. The only surviving claim about persons, rather than works, had to be rebuilt out of the law of brands.

None of this will be decided by a jury this year, or perhaps at all; cases of this kind tend to settle once a summary judgment ruling has told both sides what a trial would cost. The dates to hold are 20 October 2026, when Cisneros hears the sanctions motion; 11 November, when class certification and summary judgment motions are due; 17 February 2027, when Orrick hears them; and 1 June 2027, when, if everything holds, twelve people will be asked whether a two-gigabyte file is a copy of a hundred thousand gigabytes of art. Those who wrote this month that they already had been might have looked at the docket, where the record was waiting to be read, which is more than can be said for the weights.