Mark Lemley on GenAI Copyright
AI Initiative Speaker Series: Legal Risks with GenAI with Professor Mark Lemley is a Stanford Law School talk, uploaded June 2, 2026, in which Lemley separates generative AI copyright risk into training, output, and authorship. From the transcript, he argues current U.S. fair-use decisions favor training as transformative intermediate copying, while warning that licensing markets and unlawfully sourced books may change later cases.
The strongest contribution is his treatment of memorization and output: most generated material is not substantially similar to training inputs, but some models can reproduce heavily memorized works, plaintiff prompts can manufacture edge cases, and characters, retrieval systems, and generated code create more concrete exposure. For the site, the issue is institutional: legal responsibility moves from one author copying one work to model design, data provenance, guardrails, licensing power, and who can claim ownership of AI-made cultural labor.
The caveat is that this is a U.S.-law expert talk, not a settled map of global law or a reliability audit; Lemley himself says courts still need to confront responsibility for infringing outputs, and the rule that only humans can author copyrightable expression may come under political pressure.