Getty Images and a class of visual artists sued Stability AI in early 2023, claiming Stable Diffusion was trained on millions of copyrighted images without permission or compensation. The article examines both technical and legal terrain: Stable Diffusion learns general features rather than memorizing specific training images (supporting a non-expressive-use / fair-use defense comparable to Google Books), but can generate near-verbatim copies of training images (0.03 percent reproduction rate under adversarial testing) and synthesizes copyrighted characters and artistic styles, creating risk that fair-use claims collapse under market-substitution analysis. Legal experts assess outcome as genuinely uncertain, with potential implications: a plaintiff victory would force AI vendors to license training data from thousands or millions of copyright holders, a process so costly and slow that only large incumbents could afford frontier model training, potentially locking out smaller competitors like Stability AI and raising aggregate training costs materially. This signal touches on data provenance liability, training-data acquisition cost structure, and competitive concentration risk in model development.
420 other tracked developments touching the same claims.