The long-running battle over whether AI companies can train on copyrighted work without permission reached its first big turning points in 2026. The headline: a US federal judge approved a $1.5 billion settlement between Anthropic and a group of authors — the largest copyright class action in US history — resolving claims over books used to train its models. At the same time, courts have started drawing a crucial line: training on legally acquired material can qualify as fair use, but using pirated copies does not.
The bigger pattern is just as important. Rights holders increasingly use lawsuits to establish leverage, then convert that leverage into licensing deals. Across books, images, music, and news, 2026 looks less like a single decisive verdict and more like an industry negotiating new rules of the road, case by case and deal by deal.
The Anthropic settlement, explained
The Anthropic case is the one that made history. Authors alleged the company illegally copied their books to train its models. In a pivotal earlier ruling, the court found that training on legally obtained books was “spectacularly transformative” and qualified as fair use — but that retaining a library of pirated copies was a separate act of infringement. That distinction is the heart of the whole debate.
Rather than gamble on a trial over the piracy piece, Anthropic settled. A federal judge approved the roughly $1.5 billion deal in mid-2026, covering an estimated 500,000 works at about $3,000 per work, with a claim rate above 90%. It is the first major resolution of the book-training lawsuits and a clear signal that the downside of building on unlicensed data can be enormous.
But here’s the catch legal analysts keep stressing: a settlement is not a precedent. Because the case never reached an appeals court, it doesn’t set a binding, industry-wide rule. The core legal question — how far fair use stretches for AI training — remains genuinely open.
Fair use: a genuinely mixed record
Early rulings have gone in different directions, which is why nobody can declare a winner yet.
- On one side, the “spectacularly transformative” finding in the Anthropic matter suggested courts may view training itself sympathetically when the underlying copies were legally acquired.
- On the other, in Thomson Reuters v. Ross Intelligence, a court rejected the fair-use defense, finding that using legal headnotes to train an AI research tool was infringement.
The throughline so far: courts tend to distinguish how the data was obtained and what the model competes with. Training on lawfully accessed material for a transformative purpose fares better; ingesting pirated content, or building a tool that directly substitutes for the original, fares worse. Until an appeals court or the Supreme Court weighs in, expect this patchwork to continue.
The other big cases still in play
The Anthropic deal closes one front, but the war continues on several others:
- The New York Times v. OpenAI and Microsoft remains the most closely watched news case. More than two years in, it has not resolved, and many observers expect an eventual settlement rather than a landmark verdict — but timing is uncertain.
- Image generators are under pressure. A high-profile case brought by Disney, Universal, and Warner Bros. Discovery against Midjourney moved through discovery in 2026, testing whether outputs that resemble famous characters cross a line.
- Getty Images, which sued an AI firm back in 2023, illustrates the pivot to business: it signed a deal to bring its licensed visuals into ChatGPT’s search and discovery — sue first, license later.
Music becomes a licensing story
The music industry moved fastest from courtroom to contract. After the major labels sued AI music services Suno and Udio, the disputes began converting into deals: Warner settled and signed licensing agreements, and Universal reached a settlement with Udio tied to a new, licensed AI music platform, with Suno moving toward licensed models replacing its earlier versions. A separate multibillion-dollar suit from music publishers targeting AI training on compositions (lyrics and songwriting, not just recordings) kept the pressure on.
The lesson mirrors the rest of the field: litigation establishes leverage, licensing captures value. Rights holders would rather get paid than get an injunction — provided the price is right.
News publishers pick their lane
Publishers have split into two camps. Some, like the Times, are litigating. Others are cutting deals — Perplexity’s publisher program, for instance, shares revenue when its answer engine draws on a partner’s content. Both approaches aim at the same goal: getting compensated when AI systems ingest or surface journalism. For readers, the practical effect is a slow shift toward AI products that increasingly cite, link, or pay for the sources behind their answers. If you want to understand how these systems use and store the data you feed them, our AI Chatbot Privacy Explained: Is Your Data Safe? guide breaks it down.
What it means for you
- Your favorite AI tools aren’t going away — but they are getting more expensive to run legally, and some of that cost may eventually reach subscription prices.
- Expect more licensed content. As deals proliferate, AI outputs are more likely to draw on properly licensed images, music, and text — which can mean better quality and clearer attribution.
- Outputs that copy too closely stay risky. If a generator spits out something recognizably close to a protected character, song, or article, that’s the zone courts scrutinize most. Use these tools to create, not to clone.
- The rules are still forming. Anyone claiming the copyright question is “settled” is overstating it. The direction — toward licensing and clearer lines around piracy — is visible, but the binding law is not yet written.
You can explore the AI tools shaped by these fights in our The AI Directory.
What to watch next
Three developments would move the needle most. First, an appellate ruling on AI training and fair use — the thing that would finally create binding precedent. Second, the New York Times case, whose resolution (settlement or verdict) would set a tone for all of journalism. Third, the pace of licensing — every major deal makes the “just pay for it” model more normal and the “train on anything” era harder to defend. The trajectory of 2026 suggests a future where AI is built increasingly on licensed, accounted-for data. It just isn’t guaranteed yet.
FAQ
Is it legal for AI to train on copyrighted work?
It depends, and the law is still unsettled. Courts in 2026 have suggested that training on legally acquired material for a transformative purpose can qualify as fair use, while using pirated copies does not. But rulings have been mixed, and no appeals court has set a binding, industry-wide standard yet.
What was the Anthropic copyright settlement?
A US federal judge approved a roughly $1.5 billion settlement — the largest copyright class action in US history — over books used to train Anthropic’s models. It covered an estimated 500,000 works at about $3,000 each. Because it’s a settlement, not an appellate ruling, it doesn’t set binding precedent for the whole industry.
Did the New York Times win its case against OpenAI?
Not yet. As of 2026 the case remains unresolved after more than two years. Many observers expect an eventual settlement rather than a decisive courtroom verdict, but the timing is uncertain and no outcome should be assumed.
Are AI music generators like Suno and Udio legal now?
They’re moving toward licensed operation. After lawsuits from major labels, several disputes converted into settlements and licensing deals, with new licensed models replacing earlier versions. That doesn’t retroactively resolve every claim, but the industry is shifting from litigation toward paid licensing.
Will these lawsuits make AI tools more expensive?
Possibly, over time. Licensing content and settling claims add real costs, and companies may eventually pass some of that to users through pricing. In the near term, generous free tiers persist, but the era of building freely on unlicensed data is clearly ending.
Can I get in trouble for using AI-generated images or text?
The main legal exposure sits with the AI companies over training, not typical users. However, if a tool produces output that closely copies a protected character, song, or article and you publish it commercially, that specific output could raise infringement risk. Use AI to create original work rather than to reproduce existing content.
Zen Tech Hub may earn a commission from links on this page, at no extra cost to you.
