Andersen v. Stability AI Goes to Trial Monday: What the First AI Training-Data Case to Reach a US Jury Means for UK Rights Holders

The Case That Couldn't Be Dismissed

Andersen v. Stability AI Ltd., Case No. 3:23-cv-00201 (N.D. Cal.), was filed on 13 January 2023 by artists Sarah Andersen, Karla Ortiz, and Kelly McKernan, and it has been trying to die ever since. Stability AI, Midjourney, DeviantArt, and Runway AI moved to dismiss early and largely succeeded: DMCA Section 1202 claims went, vicarious infringement went, and the class action was substantially narrowed. What survived was direct copyright infringement. That surviving claim goes to a jury on Monday.

The docket on PACER was last updated 2 September 2026, four days before this article goes to press, confirming final pre-trial activity. This is not a drill.

What Was Actually Decided (So Far)

Judge William Orrick's August 12 ruling cleared the case for trial, finding enough factual dispute on the direct infringement claim to put it in front of twelve people who almost certainly have never heard the phrase "latent diffusion model." The plaintiffs allege that Stability AI's training pipeline ingested the LAION-5B dataset, five billion images scraped from the internet, and that the resulting model encodes their specific works in its weights. When prompted, the theory goes, the model produces infringing derivative works drawn directly from those encoded originals.

The defendants' position is the one you'd expect: training is transformative, the model learns statistical patterns rather than copying expression, and any outputs are independently created. It is the fair use argument dressed in machine-learning language, and no US court has ruled on it at trial.

The DMCA 1202 claims, which would have been the cleaner route to liability, requiring defendants to show they knowingly removed copyright management information, were dismissed. So the artists must prove the harder thing: that the model itself constitutes an infringing copy, or that its outputs do. Neither argument is easy.

Why UK Trademark Numbers Tell a Story

While American courts prepare to hear testimony about diffusion models, the UK's own market is signalling stress. AIBD analysis of IPO data via TrademarkDashboard shows 1,267 Class 45 filings in 2026 Q3, a fall of 20.4% on the prior period. Class 45 covers legal services, security services, and personal or social services, precisely the territory where AI-powered legal platforms are expanding fastest and where brand disputes between lawtech startups and established firms have already begun to surface.

That filing contraction is striking. It runs counter to the overall UK trademark trend: the IPO received 173,180 applications in 2024, a 5.8% increase on 2023, and the register now holds approximately 3 million records including around 1.1 million active registrations. When one class drops sharply against a rising aggregate, practitioners should ask why. The most plausible explanation is uncertainty: AI legal services companies are deferring trademark investment pending clearer guidance on what they can legitimately claim to offer and under whose IP they are operating.

Clearer guidance is not coming quickly.

The UK Policy Vacuum

On 18 March 2026, the UK government published its long-awaited Report on Copyright and Artificial Intelligence alongside an accompanying Impact Assessment, both required under the Data (Use and Access) Act 2025. The report's headline message was essentially: we are not legislating yet. Secretary of State Liz Kendall was explicit in her written statement to Parliament, stating the government will not rush because "it will take time to get this right."

The Lords Communications and Digital Committee had been rather less serene. Its 6 March 2026 report described the UK creative industries as facing a "clear and present danger" from generative AI trained on copyright-protected works without authorisation or remuneration. The Committee's conclusion was pointed: the problem is not UK copyright law, which it characterised as an international gold standard, but widespread unlicensed use and a lack of transparency from AI developers.

Both documents are right and both are incomplete. UK copyright law under the Copyright, Designs and Patents Act 1988 does not have a broad fair use doctrine. It has narrow permitted acts. Section 29A permits text and data mining for non-commercial research; it says nothing about commercial model training. The government's earlier preference for a commercial TDM exception with a rightsholder opt-out was withdrawn after creative industry opposition. So the statute stands as drafted, in 1988, for a world without diffusion models, and offers no clear answer.

The statute wasn't drafted for this technology, and it shows.

The Transatlantic Feedback Loop

This is why the Andersen verdict matters in London, even though the trial takes place in San Francisco.

The UK Court of Appeal is expected to hear Getty Images' appeal against the November 2025 High Court ruling that dismissed its secondary copyright claim against Stability AI. The High Court found, on narrow UK-specific grounds, that an AI model trained abroad does not infringe UK copyright when deployed here unless it stores or reproduces protected works in-jurisdiction. Getty dropped its primary infringement claim for lack of evidence. The appeal will therefore not resolve whether training itself infringes UK copyright, a gap that leaves practitioners working blind.

What Andersen may produce, if it reaches verdict, is the first judicial assessment of whether the training process produces "infringing copies" in any sense cognisable under copyright law. The US fair use doctrine differs materially from UK permitted acts; a finding that training is fair use in San Francisco does not bind Chancery Division judges. But a finding that training is infringement, that five billion scraped images lodged in model weights constitute copying, will land in the UK debate like a depth charge. Every AI developer with a training pipeline touching British creators will re-examine their exposure overnight.

In 2013, in Authors Guild v. Google, the Second Circuit held that scanning millions of books to build a search index was transformative fair use. That ruling took nine years and two appeals to settle. The AI training question is more complex by an order of magnitude, and Andersen will not settle it in a week. But it will produce the first trial record, witness testimony and expert evidence on how diffusion models actually work, and that record will shape every argument made in every court in every jurisdiction for years.

The Practical Impact

For UK AI developers, the immediate takeaway is defensive: document your training data provenance now, before discovery requests arrive. For UK rights holders, photographers, illustrators, writers, the Andersen trial is the closest thing to vindication or foreclosure that the next eighteen months offers. For in-house counsel at companies operating AI products that touch creative works, 8 September is the start of a very attentive period of reading US court filings.

The sharp contraction in Class 45 trademark filings is a lagging indicator of exactly this anxiety. Companies building AI legal services tools know they are operating in contested IP space. Some are deferring brand investment until the legal ground firms up. That is rational. It is also, frankly, a concession that the current framework is not fit for purpose.

The Nice Classification 13th Edition, which entered force in January 2026, now expressly includes notarial services in Class 45 and recognises a range of new service categories reflecting technological change. WIPO moves faster than Westminster on classification. It should not.

What Happens Next

Trial in Andersen begins 8 September 2026 before Judge Orrick in the Northern District of California. Expect it to last several weeks, with expert testimony on training data architecture as the decisive battleground. A verdict, if the jury reaches one, could come in October or November.

In the UK, the Court of Appeal hearing in Getty v. Stability AI has no confirmed date but is expected during 2026. The government's copyright reform process remains open-ended. The IPO's next quarterly trademark data will show whether Class 45 filings recover or contract further.

The jury, as it were, is literally out.