Do LLMs infringe copyright by training on other people's work?

No court has settled this yet — training AI on other people's writing, art, and journalism is the subject of dozens of active lawsuits, and the rulings so far are mixed and narrow, nowhere close to one final answer for the industry.

Why do AI companies say training is fair use?

Their argument is that learning isn't copying. An LLM doesn't store or replay the books and articles it trained on — it extracts statistical patterns from them, roughly the way a person learns to write by reading widely. On that view, the finished model isn't a copy of any single work, so training on copyrighted material to build it should count as fair, transformative use rather than infringement.

The output is something new, they argue, not a repackaging of the inputs.

Why do authors and artists say it's infringement?

Their argument is that scale and consent change everything. Creators say their work was copied wholesale, without permission or payment, to build commercial products that now compete with them — a novelist's books used to train a tool that generates novels, an illustrator's art used to train an image generator. The fact that the model doesn't paste the original straight back out, they argue, doesn't undo the unlicensed copying that happened to create it in the first place.

Why isn't there one clear answer?

Because the question sits on genuinely unsettled ground, and reasonable people read it differently. A few things keep it open:

• The law is being tested, not settled — courts are weighing old copyright principles against a technology they weren't written for.
• Jurisdictions differ — different countries can read the same facts and reach opposite conclusions, so a single global verdict is unlikely.
• The facts vary case by case — how the training data was obtained and what the model does with it can change the analysis.

Treat any confident claim that this is "obviously legal" or "obviously illegal" as oversimplified. It's a real, unresolved tension between how AI is built and who owns what it learns from — not a settled rule.