Copyright and the Training of Human Authors and Generative Machines

There are many limitations on copyright of which human authors can and do take advantage as they are learning. However, there is no blanket fair use immunity for use of copyrighted works to educate human authors, even though those authors typically do not go on to create substantially similar works. Human authors typically end up paying, directly or indirectly, for most of the copyrighted works from which they learn. Should it be different when human beings use copyrighted works to train generative AI models? This article concludes that it should not, in spite of two prominent arguments to the contrary. The first argument is that such training involves “nonexpressive use” of those works. Under the only definition of that term that distinguishes generative AI training from human learning, a “nonexpressive use” is one that does not result in an aesthetic or hedonic reaction to a work. However, copyright should be and usually has been considered to protect not just the entertainment value of works for passive and unchanging human beings, but the educational value of works for human beings who want to learn and change, both individually and collectively. The second argument is that generative AI training is functionally equivalent to human reading, viewing, or listening – activities outside the scope of copyright’s exclusive rights. However, the distinctions between and limitations on exclusive rights presuppose limited human memory and cognition, and current and future generative models are not subject to those constraints. Moreover, the very inability of computers to have any hedonic or aesthetic reactions to the works they are processing, and their inability to remember and act on those reactions, make computer processing fundamentally different than human experience of works.