Two recent court rulings in the US — involving Meta and Anthropic — offer some early but narrow guidance on how copyright law applies to generative AI. Both companies successfully argued that their use of copyrighted books to train AI models falls under fair use, but the decisions highlight the ongoing legal uncertainty around AI training data.
Here are the key takeaways:
1. Fair use applies — at least in these cases
Both Meta and Anthropic convinced federal judges that their use of copyrighted books to train their LLMs (Llama and Claude, respectively) was “transformative” enough to qualify as fair use under US law. In Anthropic’s case, Judge William Alsup noted that the method of using a scanned, purchased book database for training was sufficiently distinct from the original purpose of the books.
2. The rulings are narrow, not sweeping
While these outcomes are being celebrated by the AI industry, they don’t settle the broader debate. Both decisions were highly fact-specific and don’t establish a universal precedent. They also leave unresolved whether different training methods, data sets, or jurisdictions might lead to a different outcome.
3. Transformative use is key
In both cases, the courts focused on whether the AI companies’ use of the copyrighted material created something new and didn’t just replicate or substitute the original work. This aligns with the principle that transformative use — such as learning patterns or language from text rather than republishing it — can qualify as fair use.
4. Legal challenges are far from over
These cases only address part of the puzzle. Future lawsuits may test other questions: for example, whether outputs that closely mimic copyrighted material are infringing, or whether large-scale scraping of copyrighted content without compensation could violate other laws. Appeals in these cases are also possible.
5. The AI vs copyright debate is just beginning
While Meta and Anthropic scored wins here, these rulings highlight — rather than resolve — the tension between the rights of creators and the ambitions of AI developers. Courts, lawmakers, and regulators will likely grapple with these issues for years, and outcomes could vary significantly in other jurisdictions.