AI copyright is settling on one question: does your output compete with what you trained on?
Anthropic has agreed to pay about $1.5bn to settle a claim over books pulled from pirate libraries and to destroy the data. Meanwhile the courts are converging on a 'market-harm' test: training on someone's work is likelier to be a problem when the thing you build competes with the work you trained on, the point a legal-research AI lost on when it was trained on a rival's headnotes.
What it actually means
If you fine-tune or build a product on someone else's content, the risk isn't abstract 'was training fair' it's specific. Are you now competing with the people whose data you used? That's the line the law is drawing. There’s two practical reads from this. Where your data came from is now a product decision, not a legal afterthought, so keep the provenance of your training and fine-tuning sets like you'd keep receipts. If your AI's output substitutes for its own source, treat that as a live liability. The era of 'scrape first, argue fair use later' is closing; the cheap insurance is licensed or first-party data you can actually point to.
The legal question isn't 'was it fair to train on this' but it's 'does what you built now compete with it.'
$1.5Bwhat Anthropic agreed to pay to settle its training-data copyright case
Source: Norton Rose Fulbright · Tuesday 4 August 2026