Two AI copyright cases moved on September 8, 2026, in opposite procedural directions. Andersen v. Stability AI went to a jury before Judge William Orrick in the Northern District of California — reported as the first US jury trial over training image models on copyrighted work. NYT v. OpenAI and Microsoft went to summary-judgment argument before Judge Sidney Stein in New York, where the Justice Department filed a brief supporting OpenAI. The difference is not procedural trivia: fair use is a mixed question of law and fact, and a judge may resolve the four statutory factors alone only where the underlying facts are undisputed. A jury verdict binds nobody as precedent, applies to one model and one training corpus, and is appealable. Anthropic's $1.5 billion settlement, finally approved in July at roughly $3,000 per work, ended its case for the same structural reason: a settled case never reaches an appeals court.
Two AI copyright cases moved on Tuesday 8 September, in opposite procedural directions, and the contrast between them is more instructive than either case on its own.
In San Francisco, Andersen v. Stability AI went to trial before Judge William Orrick in the Northern District of California — reported as the first US jury trial over training an image-generation model on copyrighted work. The defendants are Stability AI, Midjourney, DeviantArt and Runway AI; the model at issue is Stable Diffusion. In New York the same day, The New York Times Company v. Microsoft and OpenAI reached summary-judgment argument before Judge Sidney Stein, over a claimed 10.8 million articles. Judge not jury.
Most coverage has labelled both landmark and moved on. The label is doing no work. What matters is why one of these disputes is being decided by twelve members of the public and the other by one judge, because the answer explains why the rulings in AI copyright have looked so contradictory, and it sets a hard ceiling on what any of them can settle.
Fair use is what lawyers call a mixed question of law and fact. The statutory test has four factors: the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of what was taken, and the effect on the market for the original. The framework is law. Most of what the framework asks about is fact.
Whether a use is commercial, how much of each work was ingested, whether the output substitutes for the original in an actual market, whether that market suffered measurable harm — these are findings about the world, established by evidence. A judge is permitted to resolve them alone only where they are genuinely undisputed. That is what a summary-judgment motion asks: not who should win the argument, but whether there is any real factual disagreement left to try. If there is, the case goes to a jury, and the jury decides those facts on the particular record in front of it.
So the split on 8 September is not two courts disagreeing about the law. It is two records in different condition. Judge Orrick had already denied motions to dismiss from Stability AI and Midjourney, finding the direct and induced infringement claims plausible enough to proceed to discovery. That case now has a factual record contested enough to require a trial. In New York, both sides are arguing that the material facts are settled enough for the judge to rule without one.
The practical consequence: a jury verdict produces something no summary-judgment opinion can. It attaches a dollar figure and a yes-or-no answer to a specific fact pattern — this corpus, this model, this market. Later defendants cannot cite it as law, but they can price against it. That is a different kind of influence from precedent, and it is the reason the trial matters commercially even though it binds nobody.
The New York case acquired a new participant this month. In September 2026 the US Department of Justice filed a brief supporting OpenAI, broadly rejecting the argument that training models on copyrighted text constitutes infringement. It is reported as the first time the federal government has taken a position in copyright litigation over the use of copyrighted material as training data.
A DOJ brief is not binding on Judge Stein and the government is not a party. But it is a named counter-argument that anyone forecasting the outcome now has to answer, and it changes the reputational arithmetic for the defendants: the position they are advancing is no longer purely self-interested advocacy. The Times argues that OpenAI and Microsoft copied its work at scale to build commercial substitutes. OpenAI argues that its models were trained on a broad and undifferentiated sweep of internet text, that they almost never reproduce the articles they read, that reported facts are not copyrightable, and that the publishers suing it have grown rather than shrunk since ChatGPT launched. Microsoft advances similar arguments for Copilot.
The most informative number in AI copyright was not produced by a court. In July 2026 a federal judge granted final approval to Anthropic's settlement of a class action brought by authors: $1.5 billion, roughly $3,000 per work across an estimated 500,000 works, over the use of pirated copies of books to train Claude. The settlement followed a ruling that training on copyrighted books can constitute fair use but that retaining pirated copies is not protected by it.
That combination is worth sitting with. Anthropic won the argument that most of the industry treats as existential — that training itself can be fair use — and paid $1.5 billion anyway, because the piracy finding exposed it to statutory damages across a class of half a million works. The lesson embedded in the figure is that how you acquired the corpus can matter more than what you did with it.
It is tempting to read the settlement as a company buying its way out of a jury. That reading is interpretation, not reporting: Anthropic did not state that as its reason, and a class action with statutory damages across 500,000 works carries enough tail risk to justify settling on arithmetic alone. What can be said without inference is structural, and it was noted at the time of approval: because the case settled, it will never reach an appeals court, so it cannot become binding precedent. A defendant who settles removes the possibility of a bad rule and also the possibility of a good one.
When the Andersen verdict lands, it will be reported as an answer to the question of whether AI training is legal. It will not be one, for four reasons that are worth holding onto in advance of the headline.
That is not a reason to ignore the outcome. It is the reason the outcome will be influential in a way that has nothing to do with law. A number attached to a fact pattern gives every general counsel in the industry something to model against, and it gives every plaintiff's firm a benchmark to open negotiations from. The Anthropic settlement already established one such anchor at roughly $3,000 a work. A jury award in Andersen would establish a second, in a different medium, arrived at by a different route.
For anyone building on or licensing generative models, the actionable read is unchanged by whichever way the verdict falls. The one finding that has already gone badly for a defendant — and cost $1.5 billion — concerned the provenance of the training data rather than the act of training. Documented, lawful acquisition of a corpus is the part of the exposure that is inside a company's control. The fair-use question is not, and will not be resolved by a single verdict in San Francisco whatever the headlines say next.
A group of artists allege that Stability AI, Midjourney, DeviantArt and Runway AI unlawfully used their copyrighted images to train or operate AI image-generation models, with Stable Diffusion the model at issue. Judge William Orrick of the Northern District of California previously denied motions to dismiss the copyright claims, finding direct and induced infringement plausible, and the case went to a jury trial beginning 8 September 2026 — reported as the first US jury trial on training image models.
We explain the mechanisms — what a ruling actually decides, what it leaves open, and who has to act on it.
More AI analysisOpenAI's new model produced two ARC-AGI-3 scores 37 points apart on the same day. The higher one cost less. What that ordering rules out, and what it means for reading any benchmark number.
Claude Fable 5.1 costs exactly what Fable 5 cost per token. The 25-to-45% saving Anthropic advertises comes from one repriced line item — cached input, now billed at 2.5% of list instead of 10%. That is a discount you only collect if you keep the agent running.
Anthropic was not breached. The malware was already on the customer's machine, and it took a session cookie rather than a password — which is why two-factor authentication did nothing and why server-side revocation is the only lever the vendor has.