Training a model involves making copies: scraping a web page, storing it in a dataset and processing it many times. Rights holders argue that copying needs a license, especially when the resulting model competes with their work. AI developers argue that training is transformative and that the model learns patterns rather than storing works. In the U.S., the main legal test is fair use, a case-by-case analysis of the purpose of the use, the nature of the work, how much was taken and the effect on the market for the original.
The U.S. Copyright Office released a pre-publication version of its report on generative AI training on May 9, 2025. It analyzes how fair use applies to training and concludes that voluntary licensing should keep developing without government intervention, with targeted measures such as extended collective licensing considered if specific markets fail. The report is guidance; courts decide the cases. In the EU, the EU AI Act requires providers of general-purpose AI models to respect machine-readable opt-outs by rights holders and publish a summary of their training content.
Companies respond by documenting where data came from. In episode 17 at 18:07, IBM’s Maryam Ashoori describes a governance process for training IBM’s Granite models that documents lineage and filters out “toxic information and copyrighted information.” That is IBM’s account of its practice; episode 17 was sponsored by Galileo. Editorially, the business takeaway is that data provenance has become a legal asset, and that vendor indemnities (see AI liability) are partly a response to this uncertainty. This entry describes the debate and is not legal advice.
Hear it from the guest
“Filter out the data, like for example, toxic information and copyrighted information”
Quotes lightly edited to remove filler words.
Sources
- U.S. Copyright Office: Copyright and Artificial Intelligence — Hub for the Office’s multi-part AI report; says the pre-publication Part 3 on generative AI training was released May 9, 2025, with no substantive changes expected in the final version; checked October 24, 2026.
- U.S. Copyright Office: Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version) — Analyzes fair use for training; concludes voluntary licensing should continue to develop without government intervention, with targeted measures such as extended collective licensing considered if market failures are shown; checked October 24, 2026.
- EU AI Act Explorer: Article 53, Obligations for Providers of General-Purpose AI Models — Requires general-purpose model providers to have a copyright-compliance policy that respects rights reservations under Article 4(3) of Directive (EU) 2019/790, and to publish a training-content summary; checked October 24, 2026.
