Top AI models from OpenAI, Google, Meta, Anthropic, and xAI can generate near-verbatim copies of bestselling novels, according to recent studies. Researchers found that models like Gemini 2.5 and Grok 3 regurgitated up to 76.8 percent of Harry Potter and the Philosopher's Stone when prompted strategically.

This discovery challenges the AI industry's core defense that large language models don't store copyrighted works. A Stanford-Yale study showed models could extract entire novels including A Game of Thrones, The Hunger Games, and The Hobbit through specific prompting techniques. Even models with safety guardrails like Claude 3.7 Sonnet could be jailbroken to produce near-verbatim text.

Legal experts warn this "memorization" ability could create significant liability for AI companies facing copyright lawsuits. A German court recently found OpenAI infringed copyright by memorizing song lyrics, while a US case involving Anthropic led to a $1.5 billion settlement after determining pirated works were "inherently, irredeemably infringing." The findings threaten the industry's "fair use" argument that training on copyrighted books transforms original works into something new.