New Study Questions AI Copyright Liability
Recent research from Stanford and Yale universities has presented groundbreaking evidence that could significantly disrupt the AI industry’s operations. Major players in the AI field, such as Google, Meta, Anthropic, and OpenAI, have long maintained that their large language models (LLMs) do not store copyrighted materials but rather learn from their training data, resembling human cognitive processes. This distinction is crucial as it forms the backbone of their defense against numerous legal challenges concerning copyright infringement.
Copyright law, particularly the U.S. Copyright Act of 1976, protects original works and grants exclusive rights to copyright owners for reproduction, adaptation, distribution, and public performance of their works. However, the industry’s argument has relied on the “fair use” doctrine, which allows limited use of copyrighted materials for purposes such as criticism and research. OpenAI’s CEO, Sam Altman, has asserted that the future of the industry hinges on its ability to utilize copyrighted data without restriction.
Despite these defenses, rights holders—including authors, journalists, and artists—have voiced strong concerns about being sidelined. They argue that AI companies train their models with their copyrighted works without fair compensation, leading to a protracted legal struggle that has already seen significant settlements.
The recent study adds weight to these concerns, revealing evidence that popular LLMs, including OpenAI’s GPT-4.1, Google’s Gemini 2.5 Pro, xAI’s Grok 3, and Anthropic’s Claude 3.7 Sonnet, are indeed capable of reproducing copyrighted texts with alarming accuracy. For instance, the study highlighted that Claude could output entire books with a 95.8% accuracy rate, while Gemini replicated passages from “Harry Potter and the Sorcerer’s Stone” and Claude produced text from George Orwell’s “1984” with over 94% accuracy. These findings challenge the assertion that LLMs primarily engage in a learning process.
The implications of these discoveries are potentially far-reaching. Copyright lawsuits could significantly escalate, costing AI companies billions in legal liabilities if the industry continues to argue that they do not replicate protected works. Legal experts, including Stanford law professor Mark Lemley, express uncertainty about whether AI models can be said to contain copyrighted texts or merely reproduce them in response to user prompts. This ambiguity complicates the ongoing discourse surrounding AI and copyright law.
Amid this, major AI companies remain committed to their position. In 2023, Google reaffirmed to the U.S. Copyright Office that no copy of the training data exists within their models, while OpenAI echoed similar claims. Critics, like The Atlantic’s Alex Reisner, label the analogy that AI learns as humans do as misleading, arguing it hinders necessary public discourse about how AI companies exploit creative works.
As the legal landscape continues to evolve, the outcomes of ongoing copyright litigation could redefine the relationship between AI technologies and intellectual property, raising critical questions about the sustainability of creative industries in an AI-driven economy.