Two copyright questions matter for AI. Is training a model on protected works lawful? And who, if anyone, owns what the model produces?
Training
United States. The answer turns on fair use, case by case. In September 2026 the Third Circuit affirmed that Ross Intelligence’s use of Westlaw headnotes to train a competing legal tool was not fair use. In other cases, courts have treated training on lawfully acquired books as fair use, while Anthropic agreed a reported $1.5 billion settlement, approved in July 2026, over books from pirate sources. The New York Times’ case against OpenAI is at summary judgment.
United Kingdom. In November 2025 the High Court largely rejected Getty Images’ claims against Stability AI, finding that the trained model was not itself an infringing copy. Getty has permission to appeal. In March 2026 the government dropped its preferred plan for a broad text and data mining exception with an opt-out, and chose to gather more evidence instead.
European Union. The text and data mining exception allows training unless rightholders opt out. Under the AI Act, providers of general-purpose models must have a copyright policy that respects opt-outs and publish a summary of their training content. In November 2025 a Munich court found OpenAI liable for reproducing song lyrics; OpenAI has appealed.
Outputs
US law requires human authorship. UK law has an unusual provision for computer-generated works, though its future is under review. In all three, the more a human selects and shapes the output, the stronger the claim.
For a business using AI, the practical risk is less about training and more about outputs that copy and contracts that do not cover it.
This article is general information on the law as at November 2026. It is not legal advice and has not been updated since that date. See the terms of use.