Copyright and AI Training Data: Where the Lawsuits Stand in 2026

Two years ago, “is training an AI model on copyrighted work legal?” was a question without a real answer — just competing theories from lawyers on both sides. In 2026, it’s a question with several actual answers, and they don’t all point the same way. More than 70 AI copyright lawsuits are active or recently resolved in US courts alone, and the first wave of real rulings and settlements has started to draw a rough map of what’s allowed, what isn’t, and what’s still genuinely unsettled. If you use AI tools to write, design, or produce anything, the outcome of these cases is quietly setting the rules you’ll eventually operate under — even though none of them are about you directly.

Thomson Reuters v. Ross Intelligence: the first real answer, and it wasn’t fair use

The case that produced the first final judgment on AI training and copyright wasn’t one of the headline-grabbing chatbot lawsuits — it was Thomson Reuters against a small legal-research startup called Ross Intelligence. Ross had trained its AI legal-search tool using Westlaw’s editorially written headnotes, and in February 2025 a federal judge ruled that this was not fair use. The decision turned heavily on market harm: Ross’s tool was built to compete directly with Westlaw, in the same market, for the same customers, using material Westlaw had invested in creating. That’s the fact pattern courts treat most harshly — not “AI trained on copyrighted text,” but “AI trained on copyrighted text in order to build a substitute for the product that text came from.” The case is now under its first appellate review, with the Third Circuit hearing oral argument in June 2026, so this rule isn’t final everywhere yet — but it’s the one existing precedent that came out clearly against an AI company.

Bartz v. Anthropic: a fair-use win, a piracy loss, and a $1.5 billion settlement

The case that’s actually shaped the most law so far is Bartz v. Anthropic. In June 2025, Judge William Alsup issued a split ruling that’s become the reference point for almost every other case since: training an AI model on copyrighted books is fair use if Anthropic legally acquired those books, but Anthropic had also built a library using pirated copies from shadow libraries, and that part of the conduct wasn’t excused by any fair-use argument about the training itself. Rather than litigate the piracy claims to a verdict, Anthropic agreed to a $1.5 billion settlement covering roughly 500,000 works — worth about $3,000 per book — with a fairness hearing held in May 2026 and final approval still under advisement as of this writing. It’s the largest reported copyright settlement in US history, and it did something more useful than any single ruling: it separated the two questions that had been getting blurred together. Training itself can be defensible. How you sourced the material to train on is a separate legal problem, and courts are treating it that way.

Kadrey v. Meta: a win for Meta that the judge himself called narrow

A similar-looking case against Meta produced a similar-looking result with a very different footnote attached. In June 2025, Judge Vince Chhabria ruled that Meta’s use of books — including some obtained from pirated sources — to train its Llama models was fair use, largely because the authors suing hadn’t presented evidence that Meta’s output was flooding the market with substitutes for their specific books. But Chhabria was explicit that this wasn’t a green light for AI training generally: he wrote that the ruling didn’t establish Meta’s conduct was lawful, only that “these plaintiffs made the wrong arguments.” That distinction matters more than the headline. Two judges, two rulings favoring AI companies, and both went out of their way to say the door is still open for a differently argued case to come out the other way.

Getty Images v. Stability AI: a loss on copyright, a narrow win on trademark

Getty’s UK case against Stability AI, decided in the UK High Court in November 2025, took a different legal route and landed somewhere unexpected. Getty argued Stability had scraped millions of its images to train Stable Diffusion; the court rejected the core copyright claim because Getty couldn’t prove the model stored or reproduced its specific images, only that it had been exposed to them during training. Getty did win a narrow trademark claim — the court found that some outputs from older Stable Diffusion versions reproduced Getty’s watermark, which is a real form of infringement, just a much smaller one than “your product was built on our photo library.” Getty has been granted leave to appeal and has a parallel US case underway in the Northern District of California. The UK case’s headline lesson: proving a model was trained on your copyrighted work is not the same as proving that work shows up, recognizably, in what the model outputs — and courts are currently requiring the second, harder thing.

What’s still unresolved

The cases most likely to set the next major precedent are still in progress. The New York Times’ suit against OpenAI and Microsoft, filed in December 2023, remains in discovery in 2026, with the Times’ side asking the court in July 2026 to sanction OpenAI over alleged discovery misconduct — a sign the case is still contentious well short of trial. On the music side, Universal Music Publishing, Concord, and ABKCO filed an amended $3.1 billion complaint against Anthropic in January 2026 alleging its training data was acquired through the same kind of piracy at issue in the Bartz case, while Sony Music’s fair-use cases against AI music generators Suno and Udio head toward a summary judgment ruling expected in the second half of 2026 — notably after Warner Music settled with Suno and Universal Music Group settled with Udio, both taking the licensing deal instead of the court fight.

The fair-use argument, in plain terms

Every one of these cases turns on the same four-factor fair-use test, but two factors are doing almost all the work: whether the use is “transformative” (training a model to generate new things is a different purpose than reading a book, which favors AI companies) and whether the use harms the market for the original work (if an AI tool becomes a substitute product that people buy instead of the original, that cuts hard against fair use). Courts so far have been fairly consistent that training itself can be transformative enough to qualify — but equally consistent that acquiring the training material through piracy, or building a tool that competes directly with the source material’s own market, breaks that protection. Neither of those principles is settled law yet; they’re patterns across a handful of trial-court rulings, several of which are now on appeal.

What this actually means if you use AI tools

None of these lawsuits are aimed at people who use ChatGPT, Midjourney, or any other AI tool to make things — they’re aimed at the companies that built and trained those tools. That’s an important distinction, and it’s why routine use of mainstream AI products isn’t the risk area. The real, practical exposure sits in a narrower set of situations: publishing AI output that closely mirrors a specific, identifiable copyrighted work (a piece of writing, art, or music that a reasonable person would recognize as derived from something particular, not just “in the style of” a broad genre); using AI-generated content commercially in a way that could plausibly substitute for the original creator’s own market, which is the exact theory that’s won in court so far; and relying on smaller or less-established AI tools whose training-data sourcing hasn’t been through the scrutiny that OpenAI, Anthropic, Meta, and Stability have now faced in court, since a tool built on clearly pirated material carries more downstream reputational and legal uncertainty than one whose provider has already litigated the point. For most everyday creative use, the risk is low and largely borne by the AI providers themselves. For commercial use that leans on a specific creator’s recognizable work, treat 2026’s court record as a warning, not a green light — the market-harm factor that’s already sunk one major case is exactly the scenario that applies.

This overview reflects publicly reported case status as of mid-July 2026; litigation moves fast and several of the cases described above have hearings, rulings, or appeals scheduled in the months ahead, so verify current status before relying on any of it. This article is general information, not legal advice — for a specific situation involving your own content or a specific AI product, consult a lawyer licensed in your jurisdiction.