Somewhere right now, a freelance illustrator is staring at a commissioned image generated in her style — her particular way of rendering light, the palette she spent a decade refining — for a client who used to pay her to make it. The client didn’t steal her files. Nobody copied a single image she made. A model simply looked at a great deal of her published work, along with millions of other images, and learned what she knows. When she says she’s been robbed, she means it literally. When the lawyers say the case is complicated, they also mean it literally. Both are correct, and the distance between those two positions is where the next decade of creative work will be negotiated.
The lawsuits now working through American courts are nominally about copying. The grievance voiced by nearly every working writer, artist and photographer caught up in them is about something else: that a machine learned from their labor and now competes with them, without paying, without crediting, without asking. Those are different claims. Only one of them reliably has a cause of action.
What the courts are actually deciding
The first rulings on training have arrived, and they are narrower than the headlines suggest. In February 2025, Judge Stephanos Bibas of the federal court in Delaware granted partial summary judgment to Thomson Reuters against Ross Intelligence, finding that Ross had copied 2,243 of Westlaw’s copyrighted headnotes to build a competing legal research tool — and that the copying was not fair use. Commercial purpose, no transformative use, direct market competition: the court’s reasoning ran against the defendant on nearly every axis. But Judge Bibas took care to note that Ross’s product was not generative AI, which is a way of saying the decision establishes less about large language models than either side would like. The ruling was certified for interlocutory appeal in May 2025 and is now before the Third Circuit, undecided.
The first federal decisions squarely on generative training came in June 2025, from two judges in the Northern District of California. In Bartz v. Anthropic and Kadrey v. Meta, both courts treated training a general-purpose model on lawfully acquired books as fair use — a genuinely significant holding for the model makers — while diverging on the treatment of pirated source material and on how much weight to give the competitive harm that AI outputs might do to authors’ markets. In Andersen v. Stability AI, a separate court let infringement claims proceed against a defendant who had merely downloaded a trained model, on the theory that the model itself might contain protectable expression. The U.S. Copyright Office, in the third part of its AI report released in pre-publication form on May 9, 2025, staked out the same conditional ground: whether a model’s weights infringe depends on whether the model has retained or memorized substantial protectable expression from the works it trained on.
What remains genuinely open, as of this writing: whether training on pirated copies can ever be fair use; whether market dilution from a flood of AI-generated outputs — works that imitate rather than copy — belongs in the fair-use analysis at all; and how much weight courts should give the fact that a licensing market for training data is now visibly forming. More rulings involving OpenAI and Google are expected in 2026. Dozens of cases are pending. There is no nationwide answer, and anyone offering one is selling something.
The gap in the middle
Here is the uncomfortable part. The thing creators are angriest about — the learned style, the distilled sensibility, the voice that now writes back at them from a chat window — is almost exactly the thing copyright does not protect. Copyright covers expression, not ideas, and style has always lived on the ideas side of that line. Human artists have absorbed each other’s styles for centuries, legally and often gratefully. The law has no mechanism for saying: yes, the model copied your books, but the actual injury is that it now paints like you.
Scholars writing in the Journal of the Copyright Society have named this a value-extraction problem — the possibility that AI systems capture the expressive value of creative work without replenishing the incentive structure that copyright exists to sustain. That framing is useful because it is honest: it is a complaint about the system’s economics, not about any particular copy. Courts can hear the copying claim. The substitution claim mostly has nowhere to go, unless a plaintiff can show the model memorized and regurgitated protectable expression, which is why so much litigation energy now goes into coaxing models into spitting out near-verbatim passages.
Meanwhile, the market is answering the question the courts haven’t. Publishers, news organizations, stock-image libraries and platforms with large archives have begun signing licensing agreements with model developers — deals with real money attached, struck while the law is still ambiguous. This is not charity and it is not capitulation; it is leverage converted into revenue while leverage lasts. But notice who holds that leverage. It is whoever owns content at scale: the conglomerate with ten million images, the publisher with a century of back issues, the platform sitting on a billion posts. The illustrator whose style opened this article does not have a licensing department. Her work is almost certainly inside some of these archives, contributing value to deals she will never see a penny from, negotiated by entities that acquired her rights in contracts she signed years ago and barely remembers.
The emerging pattern looks less like a resolution and more like a partition. The companies with something to trade get paid; the law, eventually, will draw a line around copying; and the people whose real complaint was never quite about copying will be told, accurately, that their complaint has no legal name.
One more thing is settled, though it settles nothing the creators care about. On March 2, 2026, the Supreme Court declined to hear Thaler v. Perlmutter, leaving intact the rule that copyright requires a human author. Work generated by a model belongs, in the American legal sense, to no one. Which means the machine that learned from your work produces work that cannot itself be owned — unprotected output flowing into the same market as yours, priced at the cost of electricity. The law spent three centuries building a fence around creative labor and has now, with perfect internal consistency, declined to extend the fence to the machine grazing outside it. The illustrator was told her style was never property. She is watching it become infrastructure.