Disaggregating artificial intelligence training: a transatlantic framework for copyright and generative AI

Generative AI disputes are often framed as a single question: whether training a model on copyrighted works is lawful. This article argues that the framing is misleading. AI training is a pipeline that begins with acquisition and preprocessing. It continues through intermediate copying during optimisation, retention and deployment, and each stage raises distinct copyright questions. Recent US decisions diverge in result but share a stage-specific method. Bartz v Anthropic separated transformative optimisation from a pirated library. Kadrey v Meta made market harm turn on evidence. Thomson Reuters v Ross Intelligence rejected fair use where copied editorial content supported a directly competing product. The English High Court’s Getty Images judgment likewise distinguished upstream copying, the legal character of model weights and downstream outputs. Together they support analysis through provenance, purpose and substitution. This article develops that comparative US–UK framework, situates it within the DSM Directive and EU AI Act, and translates it into proposals for pleading, discovery and documentation. It also develops proposals for retention, output controls and remedies. A South Korean survey serves only as contextual evidence of bounded compensation preferences. The central claim is methodological: liability should attach to the relevant operation, copy, actor and territory. The analysis should separately address the resulting market effect rather than ‘training’ in the aggregate.