Copyright boundaries of training data: a diligence checklist for global AI teams
Where data comes from, how far the license goes, and whether outputs are substantially similar — three questions frame the legal risk of training data.
Training-data copyright is among the most uncertain and most overlooked parts of taking AI global. Jurisdictions differ sharply on 'fair use' and text-and-data-mining exceptions.
The first question is provenance: scraped public data, licensed datasets, and user-uploaded content have entirely different boundaries, and each source and license scope must be mapped.
The second is the license chain: whether a dataset's sub-licensing covers training use, allows commercial use, or carries attribution or share-alike terms — often buried in long license agreements.
The third is output: when a model's output is substantially similar to training material, risk shifts from 'training' to 'generation'. Provenance records and filtering are the sustainable approach.