Lemley & Cooper paper: LLMs may not contain 'copies' under copyright law
New legal analysis argues extractability determines if LLMs infringe copyrights.
Recent research shows it's possible to extract verbatim or near-verbatim text of some copyrighted works from large language models, indicating that models 'memorize' training data. However, LLMs don't store information like databases; their weights encode probabilistic token relationships. A new paper by Mark A. Lemley and A. Feder Cooper, forthcoming in the Berkeley Technology Law Journal, tackles a critical legal question: does a probabilistic representation that might generate a copyrighted work constitute a 'copy' under copyright law? The authors note that existing statute and case law are largely unhelpful.
The paper argues that courts will likely adopt a functional approach: an LLM contains a copy only if the work is straightforward to extract from its outputs. This standard is unsatisfying as a policy matter, because it creates legal uncertainty and may fail to protect creators whose works are only occasionally regenerated. The authors suggest potential changes to copyright law, such as redefining what constitutes a copy in the context of probabilistic models. The ruling has massive implications: if extraction difficulty is the test, many LLMs could avoid infringement liability even if their training data included copyrighted works without permission.
- LLMs store probabilistic token relationships, not exact copies, complicating copyright analysis.
- The paper predicts courts will rule a 'copy' exists only if extraction from outputs is straightforward.
- Authors propose legal reforms because the functional approach is unsatisfying and leaves policy gaps.
Why It Matters
This legal analysis could shape whether entire classes of LLMs are deemed infringing, affecting AI regulation and training data practices.