The legal characterization of high-dimensional neural network weights, embedding vectors, and diffusion model parameter checkpoints as potential derivative works.
At the intersection of machine learning linear algebra and Title 17, courts must resolve whether billions of floating-point weights in a foundation model constitute an unauthorized 'recast, transformed, or adapted' derivative work under § 101. Plaintiffs argue that training captures the expressive essence of protected works within compressed mathematical vectors. Defendants maintain that weights contain uncopyrightable statistical correlations, facts, and generalized visual concepts, rather than protectable expressive fragments.
Northern District of California decisions (Andersen v. Stability AI, Tremblay v. OpenAI) have largely dismissed derivative work claims absent plausible allegations that the model outputs are substantially similar to specific training works, while leaving direct infringement and DMCA § 1202 claims alive.
