Foundation Models and Copyright

Foundation models use works on a scale and for a purpose that strain the distinction between protected expression and non-appropriable information. The article rejects binary answers on the lawfulness of training and argues that the correct legal unit of analysis is not the model but the chain of technical acts that produce and exploit it: acquisition, copying, preparation, training, retention, fine-tuning, generation and communication. At each link the acts, the actors and the defences change. Starting from Argentine Act 11,723, which contains no general text and data mining exception, and comparing the European Union, the United Kingdom, Germany and the United States, it proposes a stage-by-stage matrix and identifies three cross-cutting requirements: lawful access, corpus transparency and examination of market effects. The 2025 and 2026 case law is used to illustrate the links in the chain, not as law applicable in Argentina. English translation of the author's Spanish-language article; preprint, not peer reviewed. Keywords: foundation models; copyright; training; text and data mining; corpus transparency; memorisation.