Research Reveals Meta AI Model Can Nearly Replicate Half of a Harry Potter Book

The legal challenges Meta faces regarding copyright infringement differ from those encountered by Google Books due to its availability of downloadable content. While Meta could argue that using 42% of Harry Potter falls under fair use, the analysis is complex, involving both the training set and model incorporation. Researchers from Cornell and Stanford accessed necessary model components, highlighting risks for open-weight models compared to closed-weight ones. Leading labs are now restricting access to critical data, complicating external research on models like Llama 3.1 70B. This may discourage companies from releasing open-weight models, as maintaining servers enables filtering of potentially infringing outputs. Although this situation might deter openness, judges could also view the public sharing of model weights favorably, thus possibly lessening skepticism towards companies like Meta. Overall, the tension between copyright law and the need for transparency in AI development is evident.

Source link