Traditional cultural expressions used in generative-AI training raise distinct questions about copyright in recorded expression, lawfulness of acquiring training copies and the terms on which archives publish and license collections. This article argues that institutional release decisions are an important but incomplete governance lever, not a substitute for intellectual-property rights or a legal prohibition on independent scraping. The 2024 WIPO Treaty on Intellectual Property, Genetic Resources and Associated Traditional Knowledge concerns patent disclosure rather than AI training. In Bartz v Anthropic and Kadrey v Meta, United States district courts reached fact-specific fair-use holdings concerning literary training data; neither judgment creates a general immunity for training on traditional expressions. Bartz also distinguished training use from a retained pirated-book library, although that litigation subsequently ended in a court-approved settlement and dismissal that did not decide the merits. New Zealand’s Te Papa disclosed established Māori consultation and collection-access practices, alongside reported absences of AI-specific risk assessments and contract clauses. Nepal’s Copyright Act expressly excludes folk songs, folklore and proverbs as such from copyright while protecting some original derivative expressions. The article proposes differentiated consent, contract, technical and statutory safeguards, with attention to their enforceability and limits.