Training data transparency is the principle, and increasingly the legal obligation, that developers of AI models disclose what content was used to train them, where it came from and how it was processed. In copyright & AI, it is the practical precondition for every other rule: a rightsholder cannot enforce an opt-out, negotiate a license or bring an infringement claim without knowing whether their works entered a training set.
The best-known binding version is the public training content summary that the EU AI Act requires from providers of general-purpose AI models. Similar disclosure duties exist in some US state laws, and voluntary documentation practices such as datasheets and model cards predate the legislation by several years.
What training data transparency requires in the EU
Under the AI Act, every provider of a general-purpose AI model must draw up and make publicly available a sufficiently detailed summary of the content used for training, following a template issued by the AI Office. The obligation has applied since 2 August 2025 for new models, while models already on the market before that date have until 2 August 2027 to comply.
The summary is deliberately not a full list of works. It must be comprehensive in scope rather than technically detailed, enough for rightsholders to exercise their rights while allowing providers to protect trade secrets. The table shows the main parts of the template.
| Section | What the provider discloses | Purpose for rightsholders |
|---|---|---|
| General information | Model identity, modalities, approximate data size | Understanding the scale of training |
| Public datasets | Named large public datasets used | Checking known collections |
| Licensed private data | Categories of licensed or purchased data | Seeing which sources were paid for |
| Web-crawled data | Crawlers used, period, most relevant domain names | Detecting scraping of own sites |
| User data | Whether user interactions were used | Assessing data from own services |
| Processing aspects | Opt-out compliance, removal of illegal content | Verifying respect for reservations |
For web-crawled content, the template asks for a list of the most significant domain names, measured by the share of crawled data, with lighter thresholds for smaller providers. Alongside the summary, providers must keep a copyright compliance policy that respects machine-readable reservations under the text and data mining exception.
Levels of disclosure
Transparency is a spectrum, and arguments about it usually concern how granular disclosure should be. Rightsholder groups typically ask for work-level information, while developers argue that lists of billions of items are neither feasible nor useful and would expose commercially sensitive choices.
| Level | Typical content | Main trade-off |
|---|---|---|
| Category level | Types of sources, such as books or code | Easy to publish, hard to act on |
| Source level | Named datasets and top domains | Balance between utility and secrecy |
| Dataset documentation | Datasheets with collection and filtering methods | Useful for researchers, costly to maintain |
| Work level | Searchable list of individual works | Most actionable, rarely feasible at scale |
| Full data access | Inspection of the dataset itself | Usually limited to regulators or courts |
The EU template sits at the source level. Work-level confirmation usually requires either a court order for disclosure, a regulator request or technical methods such as membership inference, which remain unreliable for individual items.
How it relates to other copyright and AI rules
Training data transparency does not decide whether training is lawful; it makes that question testable. Its value depends on the substantive rules around it, which differ sharply between jurisdictions.
| Concept | Question it answers | Link to transparency |
|---|---|---|
| Text and data mining exception | When is copying for analysis allowed | Disclosure shows whether the exception was relied on |
| TDM opt-out | Can a rightsholder reserve rights | Summaries reveal whether opt-outs were honored |
| Fair use | Is unlicensed use excused in the US | Discovery in litigation substitutes for disclosure |
| AI training licensing | Who was paid for which content | Licensed sources appear in the summary |
| Deepfake labelling | Is output marked as synthetic | Covers outputs, not training inputs |
In the United States there is no federal disclosure duty. California AB 2013 requires developers of generative AI systems offered to Californians to post high-level documentation of training datasets from January 2026, while most other detail emerges through litigation discovery in copyright cases.
Compliance in practice
Meeting transparency duties starts long before a summary is written. Providers need data provenance records from the moment of collection: which crawler fetched which domain and when, which robots.txt and metadata reservations were observed, and which licensed corpora entered which training run.
- Log crawling at the domain level, including dates and the user agent, so the top-domain list can be generated reliably.
- Record opt-out handling, such as robots.txt rules and machine-readable reservation protocols, together with the date each rule was read.
- Version the summary with each new model release or significant retraining.
- Align the summary with the copyright policy, because inconsistencies between the two invite scrutiny.
- Prepare for enforcement: the AI Office can request information, and fines for providers can reach 3 percent of global turnover or 15 million euros, whichever is higher.
Limitations and criticism
Critics on both sides see weaknesses. Rightsholders point out that a domain list does not reveal whether a specific work was used, so the summary narrows the search without settling it. Developers warn that disclosure may expose data strategies and increase litigation risk without improving model quality.
There are also verification gaps. A summary is self-declared, and checking it against the actual dataset requires access that outsiders lack. Open-weight releases add another nuance: some open-source exemptions in the AI Act do not cover the training content summary, so even freely licensed models must publish one.
In everyday language
Training data transparency is like an ingredients label for an AI model. A food label does not list every grain of flour, but it tells you what kinds of ingredients went in and where the main ones came from.
In the same way, AI developers must publish an overview of what they trained on. Writers, artists and publishers use it to see whether their work was likely included, and then decide whether to object, ask for payment or check that their opt-out was respected.
Seen in the real world
A mid-sized European news publisher has reserved its articles against text and data mining since 2023 through robots.txt rules and a machine-readable reservation in its site metadata. When a new general-purpose model is released, the publisher’s legal team opens the provider’s published training content summary.
The summary lists crawling periods and the most significant domains, and the publisher’s domain appears among them with a crawl window that started before the reservation was added. The team compares server logs with the stated crawler names and confirms the visits. Using the summary and the logs as a basis, the publisher asks the provider to explain how its reservation was handled and opens license negotiations for future model versions, rather than going straight to court.