On 29 July 2026, the China National Intellectual Property Administration (CNIPA) issued supplementary examination guidance for large language models and generative AI inventions, formally bringing training-data compliance into the patent review process. Where an AI model has been pre-trained or fine-tuned on copyright-protected datasets, applicants are expected to include a declaration addressing the legality of the data source. The guidance also introduces an initial “data source whitelist” covering several trusted open-source datasets; applications that clearly rely on listed datasets may avoid an additional copyright review and qualify for a priority examination route.
The practical effect is to move data provenance from a back-office governance issue into the drafting and filing strategy itself. The whitelist could reduce friction, but its value will depend on scope, update frequency and day-to-day examination practice. Applicants should therefore preserve licence terms, dataset versions, acquisition records and internal use logs before filing, rather than trying to reconstruct the evidence during prosecution.
