Who Owns the Internet AI Trained On

In 2023, authors began discovering that their own books had trained the chatbots now answering millions of questions a day. Within three years, the same underlying question was in front of federal judges in California, Delaware and New York: is training an AI model on copyrighted writing fair use, or copying at an industrial scale copyright law was never built to allow?

This video explains what training actually does to a book (statistical weights, not a stored copy), the four-factor fair-use test every one of these cases turns on, and the shadow-library piracy allegations - LibGen and Books3 - common to the cases against Anthropic, Meta and OpenAI.

It covers Thomson Reuters v. Ross Intelligence, the first case to reach a final appellate ruling: Judge Stephanos Bibas's February 2025 summary judgment and the Third Circuit's September 2026 affirmance, which limited its reasoning to Ross's non-generative tool. It covers Kadrey v. Meta, where Judge Vince Chhabria granted Meta a narrow win on a thin record, warning in his own words that the decision 'does not mean that Meta's use of copyrighted materials to form its language models is legal.' And it covers Bartz v. Anthropic, where Judge William Alsup split his ruling in two - training on legally acquired books was fair use, training on pirated copies was not, writing that Anthropic offered no justification 'except for Anthropic's pocketbook and convenience' - a finding that led to a proposed $1.5 billion settlement, the largest copyright class-action settlement on record.

It explains The New York Times' case against OpenAI and Microsoft, including the Times' June 2026 amended complaint narrowing its claims and sharpening its case against Microsoft's own infrastructure. And it covers the consolidated Authors Guild multidistrict litigation before Judge Sidney H. Stein, which survived a motion to dismiss in October 2025 and reached cross-motions for summary judgment in September 2026.

The video closes by showing why these rulings do not line up into one rule: courts have been separating training, acquisition and output into three different legal questions, answering one or two at a time, case by case - which is exactly why there is still no single settled answer to who owns what an AI model learned from.

Educational documentary. Not financial or investment advice.