51 copyright lawsuits have been filed against AI companies but only 3 judges have ruled on fair use (2 for, 1 against), leaving the $12.3B training data economy in legal limbo

As of October 2025, 51 copyright lawsuits have been filed against AI companies including OpenAI, Meta, Stability AI, Anthropic, and Google over the use of copyrighted works to train AI models, but only 3 federal judges have issued rulings on the core fair use question -- 2 ruled in favor of AI companies (including Meta's June 2025 summary judgment victory) and 1 ruled against. Meanwhile, Anthropic settled a book-training lawsuit for up to $1.5 billion after a court found it may have illegally downloaded 7 million books. The New York Times has demanded 20 million private ChatGPT conversations in discovery against OpenAI. Why it matters: The legal uncertainty around training data means no AI company knows whether its foundational models are built on lawful inputs, so companies cannot accurately assess litigation risk or set aside appropriate reserves, so content creators from individual authors to the New York Times cannot enforce their rights or negotiate fair licensing terms, so a de facto regime of 'train first, litigate later' rewards companies with the deepest legal war chests, so smaller AI startups and researchers who cannot afford billion-dollar settlement risk are deterred from building foundation models, concentrating the industry further. The structural root cause is that US copyright law's four-factor fair use test was designed for human-scale reproduction and transformation, not for the ingestion of entire corpora of human knowledge into statistical models, and Congress has not passed any legislation clarifying whether machine learning training constitutes fair use -- leaving the question to be resolved case-by-case across dozens of federal courts with no binding precedent.