Data, Copyright and Content Licensing

Every large AI model is built from other people's work: books, news, code, photographs, songs, forum posts and the open web. For most of the deep-learning era nobody asked permission or paid. Since late 2022 that has produced the largest body of copyright litigation the technology industry has seen: about fifty US suits by September 2025 (tracker count, verify) plus test cases in the UK, Germany, India and Canada. The first rulings, in 2025, split the problem in two. Training on lawfully acquired works was held to be fair use in two San Francisco cases (Bartz v. Anthropic and Kadrey v. Meta, June 2025). Acquiring works from pirate libraries was not excused, and that produced the largest copyright settlement in US history: $1.5 billion from Anthropic, about $3,000 per work on a list of ~482,000 works, finally approved in July 2026 (reported). A non-generative competitor lost (Thomson Reuters v. Ross, affirmed on appeal September 2026). A German court held that a model which memorised lyrics contained copies of them (GEMA v. OpenAI); an English court held that Stable Diffusion's weights did not (Getty v. Stability AI), both in November 2025. Alongside the suits a licensing market has grown: Reddit with Google and OpenAI, News Corp with OpenAI (reported at more than $250 million over five years), Amazon with The New York Times (reported at $20–25 million a year), Disney's $1 billion investment in OpenAI with a character licence for Sora (December 2025), and record-label suits against Udio, and Warner's against Suno, turned into licences (UMG and Sony filed a second suit against Suno on 18 September 2026, reported). This section explains the law in business terms (why acquisition and outputs matter more than training), the main cases and their status at November 2026, the licensing market and its pricing, opt-outs and technical controls, the EU, UK and Asian regimes, and how rights risk enters a lab's or app's economics. A worked example compares licensing with litigation exposure, and a checklist sets out provenance questions.