Ask five people what the AI copyright cases have decided and you will get five different answers, most of them wrong in the same direction: that a verdict is coming, or has come, that will settle whether training a model on scraped creative work is legal. It hasn't happened. The four cases people usually mean when they say "the AI copyright cases" are sitting in four different places, on four different questions, and not one of them has produced a ruling on the question a working creator actually cares about: is training on your work, without your permission, infringement. Here is where each one actually stands as of this week, and why waiting on any of them is not a plan.
Andersen v. Stability AI: the one built to answer the question, still not decided
Andersen v. Stability AI Ltd., filed by artists Sarah Andersen, Kelly McKernan, and Karla Ortiz in January 2023, is the closest thing to a direct test of the core question. The surviving claims argue that a diffusion model's weights embody copies of the training images, which would make training and distributing the model itself the infringing act, not just anything a user later generates with it. That is the theory a jury would actually have to weigh.
No jury has weighed it yet. The case was originally on track for a September 2026 trial date; the docket has since moved, and jury trial is now scheduled for 5 April 2027 before Judge William Orrick in San Francisco, with a pretrial conference set for 1 March 2027. Nothing about liability has been decided. Any claim that this case has already gone to a jury, or produced a verdict, is describing a date that no longer applies.
New York Times v. OpenAI and Microsoft: still fighting over summary judgment
The Times' suit against OpenAI and Microsoft turns on a different fact pattern: alleged verbatim or near-verbatim reproduction of paywalled articles in chatbot output, not just training on the underlying text. A federal judge in the Southern District of New York let the core claims proceed in 2025, and the case has stayed in the pretrial fight since. On 1 September 2026, the U.S. Department of Justice filed a Statement of Interest urging the court toward a training-stage fair-use finding while treating output-level copying as a separate question. On 17 September 2026, an unredacted version of the Times' own filing was unsealed as part of its motion for summary judgment, putting internal company communications about content use into the public record.
None of that is a ruling on the merits. Summary judgment motions are still being briefed. Whatever the court eventually holds will turn on a specific record of specific outputs allegedly copying specific New York Times articles, which is not a fact pattern most individual creators can borrow.
Getty Images v. Stability AI: the one that actually finished, and what it didn't decide
This is the case people most often mistake for a green light or a red light on AI training generally. It is neither. The England and Wales High Court handed down judgment on 4 November 2025, and Getty's copyright claim against Stability AI failed. But it failed on narrow, procedural grounds: Getty could not show that Stable Diffusion's training and development actually took place in the UK, and dropped its primary copyright and database-right claims before closing submissions rather than have them decided against it on the facts. The court did find limited trade mark infringement, holding Stability responsible for marks appearing in some model outputs. The judgment says almost nothing about whether training a model on copyrighted images is itself infringing, because that question was never properly put to the court. Treating this as a decided answer on AI training and copyright, in either direction, overstates what actually happened.
Bartz v. Anthropic: resolved by a settlement, not a ruling
The largest of the four in dollar terms is also the one furthest from answering the legal question. Judge Araceli Martínez-Olguín granted final approval to a $1.5 billion class settlement on 20 July 2026, closing the case with a 91.3% claims rate among the class and roughly 350 valid opt-outs. That is a negotiated payment to end litigation risk, not a court holding on whether training a language model on pirated books is fair use. Anthropic settled specifically over the alleged use of pirated copies to build a training dataset, a narrower and more damaging fact pattern than "trained on the open internet." Nothing in the settlement binds any other company, any other dataset, or any other court.
What none of these decide for you
Even a decisive ruling in any one of these cases would not hand any other creator a ready-made answer. Each turns on its own record: which specific works were used, whether the defendant's own conduct (pirated sourcing, verbatim output, in-jurisdiction training) matches the theory being tested, and whether the plaintiffs can actually prove they made the work first and controlled it since. Courts decide the case in front of them, on the evidence in front of them. A win for Andersen's class would not automatically cover an illustrator who isn't in it. A loss would not automatically doom one either. What carries across every one of these cases, win or lose, is that the plaintiffs who got furthest were the ones who could show, with evidence, what they made and when.
What you can control while the cases run
You do not control when Andersen goes to trial or how the Times' summary judgment motion comes out. You do control whether you can prove, today, that a given piece of work existed in your hands at a given moment, unchanged since. Under eIDAS Article 41, a qualified electronic timestamp carries a legal presumption of the accuracy of its date and time and of the integrity of the data it covers, in every EU member state. Copyright itself is automatic under the Berne Convention, which needs no registration in any of its 181 member states, but automatic protection still leaves the burden of proof on the person claiming authorship the moment a dispute starts. A sealed, timestamped record of your own work does not depend on how any of these four cases turns out, and it is not waiting on a court calendar that has already slipped once. It is proof you hold before you ever need it.
If you make visual or written work that AI systems could plausibly train on, the practical move is not to track four separate dockets and hope one resolves your specific question. It is to build your own evidentiary record as you create, so that whenever the question does reach you, you are not starting from nothing. See how that works for creators specifically at /solutions/creators.
