AI training data and copyright, explained

The training fight is about the tool's builders, not its users. But the answers could change what tools cost, what they can do and what terms you agree to.

What the argument is

• Models learn from huge collections of existing work. Building training sets can involve "making copies of copyrighted works". That raises a question. Does it infringe the owner's right to make copies, or does it fall under fair use?¹

• Rights holders and governments are asking for clarity. The UK government says "the widespread use of copyright material for training AI models has presented new challenges". It wants "ensuring AI developers provide right holders with greater clarity about how they use their material".²

• Developers say training is allowed. Wikipedia says US developers "have traditionally believed" training to be allowable under fair use. Whether it is, is what the courts are testing.¹

Cases worth knowing about

• Getty Images v. Stability AI. Getty sued in London in January 2023 over use of its images in training data without a licence.¹

• A code case. A 2022 class action said an AI coding tool had "violated the copyright of the repositories' authors".¹

• Rulings so far are mixed. In the Anthropic case, the judge held that Anthropic's "use of purchased books for training was within fair use", but did not extend that to the Pile with unlicensed works. In a case against Meta, the judge stressed his ruling did not cover every use of copyrighted material. In Thomson Reuters v. Ross, a non-generative tool, unlicensed use for training was held not to be fair use.¹ These are Wikipedia's summaries, not the rulings themselves.

These are summaries from Wikipedia, and cases move on. Check a current, reliable report before you quote any of them.

What governments are doing

See what the official copyright offices say for more.

Does it touch your business?

• Not directly, for now. The suits above name the builders of the models, not the businesses that use them.

• Your contract may shift risk to you. OpenAI's terms say "If you are a business or organization", you cover it against third-party claims tied to your use, as far as the law permits. Look for the clause in your tool's terms.⁵

• Tools and prices may change. If courts or lawmakers set new rules, tools may change what they train on or how they license it.

• Ask your supplier what they say about training data. Keep any written answer with your records.

• Don't repeat unsettled claims. Avoid telling customers a tool is "clean" or "licensed" unless the maker states it in writing.

Nothing here is advice for your own case. A lawyer can give that.