What rules require AI companies to disclose their training data?
The EU AI Act requires every provider of a general-purpose AI model to publish a "sufficiently detailed summary" of the content used for training, in a template the European Commission had published by 1 August 2025. In the US, no law requires a summary; the Copyright Office's May 2025 report on training says licensing markets are growing and government intervention would be premature.
What could your company's records earn? Get a range in 30 seconds. Free, no obligation.
What does the EU AI Act require about training data?
Article 53(1)(d) of the EU AI Act requires providers of general-purpose AI models to make public a "sufficiently detailed summary" of the content used for training, following a template from the AI Office (AI Act Service Desk, Article 53). The Commission says using its template is mandatory.
| Item | What the Commission's page says | Date |
|---|---|---|
| Who must publish | Any provider that places a general-purpose model on the EU market, including open-source models | EU Commission template FAQ, read 21 Nov 2026 |
| Start date | 2 Aug 2025 for new models | EU Commission, 1 Aug 2025 |
| Older models | Summary due by 2 Aug 2027 for models on the market before 2 Aug 2025 | EU Commission FAQ |
| Updates | At six-month intervals, or sooner if new training data changes the summary materially | EU Commission FAQ |
| Enforcement | AI Office enforcement powers start 2 Aug 2026 | AI Act Service Desk |
| Fine | Up to 3% of worldwide annual turnover or €15,000,000, whichever is higher | EU Commission FAQ |
What must the training summary list?
The template has sections for the sources of training data. It names publicly available datasets, private datasets, data scraped from online sources, user data and synthetic data. For scraped content it requires the names of crawlers, the collection period, a description of what was scraped, and a list of the top 10% of domains scraped (for small and medium companies, the top 5% or 1,000 domains, whichever is lower).
For user data, the template requires disclosure of the kinds of data and the services they came from. It does not require disclosure of any personal information (EU Commission FAQ on the template).
What does this mean for a company licensing its records?
Records licensed from a company would most likely fall under the template's "private datasets". The AI provider has to describe such datasets in its public summary. Buyers who sell models in the EU therefore have a reason to want clean records of where each dataset came from. That is one reason buyers ask for proof you own the records and written permission to license them. See what AI data buyers check and data license contract clauses.
What has the US Copyright Office said about AI training?
The Copyright Office is issuing a report called "Copyright and Artificial Intelligence" in several parts. It opened a public inquiry in August 2023 and had received over 10,000 comments by December 2023 (Copyright Office, AI page).
| Part | Topic | Date |
|---|---|---|
| Part 1 | Digital replicas | 31 Jul 2024 |
| Part 2 | Copyrightability of outputs made with generative AI | 29 Jan 2025 |
| Part 3 | Generative AI training (pre-publication version) | 9 May 2025 |
Part 3 reaches three conclusions. Several stages of building a model use copyrighted works in ways that implicate owners' rights, and whether those uses are fair use depends on what works were used, from what source, for what purpose and with what controls on outputs. Licensing agreements for AI training, individual and collective, are emerging quickly in some sectors, though their availability is inconsistent. And because voluntary licensing is growing and stakeholders did not back a statutory change, the Office says government intervention would be premature (Copyright Office, Part 3, May 2025).
Where do official filings show AI data revenue?
Company disclosures give the best price evidence. Reddit's 2024 S-1 reported licensing arrangements with a $203.0 million total contract value and terms of two to three years (SEC filing, 22 Feb 2024, as quoted by TechCrunch). A federal bankruptcy judge approved Google's $10 million purchase of Spirit Airlines' data on 17 Nov 2026 (Reuters). For all public deals, see the AI data licensing deals tracker.
Is there a US rule that matches the EU summary?
In the pages we read, no. The Copyright Office report is advice to Congress and says it cannot prejudge the result in any particular case. It is not a law, and it sets no disclosure duty. State privacy laws and contract terms still shape what you can license. See AI data licensing explained.
Questions
Does the EU summary rule apply to US companies?
It applies to any provider that places a general-purpose AI model on the EU market (EU Commission FAQ). That can include providers based in the US.
Will my company's records be named in a public summary?
The template asks for descriptions of private datasets, not a copy of the data. It does not require personal information to be disclosed.
Is the Copyright Office report final?
Part 3 is a pre-publication version. The Office says a final version will follow with no substantive changes expected to its analysis or conclusions.
Does licensing remove copyright risk?
A license gives the buyer permission for the records you own and may license. It does not cover records you do not own.
Sources
- EU AI Act, Article 53, via the Commission's AI Act Service Desk, read 21 Nov 2026: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-53
- European Commission, "Template for general-purpose AI model providers to summarise their training content", read 21 Nov 2026: https://digital-strategy.ec.europa.eu/en/faqs/template-general-purpose-ai-model-providers-summarise-their-training-content
- European Commission, "EU rules on general-purpose AI models start to apply", 1 Aug 2025: https://digital-strategy.ec.europa.eu/en/news/eu-rules-general-purpose-ai-models-start-apply-bringing-more-transparency-safety-and-accountability
- AI Act Service Desk, enforcement powers from 2 Aug 2026: https://ai-act-service-desk.ec.europa.eu/en/ai-act/faq/commissions-enforcement-powers-related-ai-act-obligations-providers-most-advanced-models
- U.S. Copyright Office, "Copyright and Artificial Intelligence", read 21 Nov 2026: https://copyright.gov/ai/
- U.S. Copyright Office, "Copyright and Artificial Intelligence, Part 3: Generative AI Training (Pre-Publication Version)", May 2025: https://copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf
- TechCrunch, "Reddit says it's made $203M so far licensing its data", 22 Feb 2024: https://techcrunch.com/2024/02/22/reddit-says-its-made-203m-so-far-licensing-its-data/
- Reuters, "US bankruptcy judge approves Spirit Airlines data sale to Google", 17 Nov 2026: https://www.investing.com/news/stock-market-news/us-bankruptcy-judge-approves-spirit-airlines-data-sale-to-google-5034262