Market dilution, fair use, and generative AI

To build foundation models, neural networks must feed on vast amounts of data, including many copyrighted works. If courts decide that training generative-AI models on copyright-protected, human-authored works is illegal, liability risks for all developers and users of these models would increase dramatically, although noncommercial developers and users would be less vulnerable than commercial developers and users. AI developers, including many scientific laboratories, have been counting on the “fair use” doctrine of US copyright law to avoid liability for training on these works. Fair use is intended to promote freedom of expression and creativity by allowing unlicensed use of copyrighted works in certain situations—including, in many cases, research, teaching, and commentary, but also more broadly where enforcing copyright protections would stifle rather than advance copyright’s goal of encouraging the production of new works. So far, AI developers’ reliance on fair use has been mostly vindicated. Two federal courts have ruled that using copyrighted works to train generative-AI models is fair use. But the fight is far from over.