Ars Technica AI

AIs Can Generate Near-Verbatim Copies of Novels From Training Data

Recent Stanford and Yale studies show that leading large language models—including OpenAI's, Google’s Gemini 2.5, Anthropic’s Claude 3.7, and xAI’s Grok 3—can generate near‑verbatim excerpts from copyrighted novels, reproducing up to 77% of the original text when prompted. The findings reveal that LLMs memorize far more of their training data than the industry has publicly acknowledged, contradicting claims that models only learn abstract patterns. This memorization challenges the legal defense of “fair use” and raises questions about the adequacy of existing model safeguards. The issue extends beyond literature, posing privacy and liability concerns for sectors such as healthcare and education.