Mathematician argues AI training mathematically precludes copyright infringement

A mathematician posted a detailed argument on Reddit asserting that generative adversarial network training processes mathematically preclude copyright infringement by preventing narrow derivativeness. The author contends that because GANs utilize vast datasets where neither generator nor discriminator can perfectly memorize individual inputs, outputs cannot be narrowly derivative of specific copyrighted sources. This technical defense challenges prevailing legal theories that equate AI training with unauthorized reproduction. The post defines narrow derivativeness as dependence on few sources and argues proper training runs inherently violate this condition through backpropagation mechanics. While acknowledging non-legal expertise, the author presents this as a mathematical proof rather than mere opinion. Critics maintain that substantial similarity tests remain applicable regardless of underlying architecture. This debate highlights growing tension between computer science explanations of machine learning and traditional copyright frameworks currently being tested in federal courts.