# The EU AI Act and training data transparency

URL: https://sellmydatatoai.net/guides/eu-ai-act-training-data-transparency
Reviewed: 2026-11-08

Article 53(1)(d) of the EU AI Act requires every provider of a general-purpose AI model to publish a sufficiently detailed summary of its training content, using the template the Commission published on July 24, 2025. The duty has applied since August 2, 2025, Commission fines of up to 3% of worldwide turnover apply from August 2, 2026, and older models have until August 2, 2027.

## Key takeaways

- Every general-purpose AI model provider must publish a training content summary, including providers of open-source models.
- The summary lists large public datasets and the top 10% of scraped domains, but licensed datasets need only a yes or no answer and the modality.
- Commission enforcement and fines began on August 2, 2026, and models placed on the market before August 2, 2025 must comply by August 2, 2027.
- The Digital Omnibus on AI, Regulation (EU) 2026/1744, delayed high-risk rules but left the Article 53 training data duties unchanged.

The EU AI Act is the first law to require AI model providers to say publicly what their models were trained on. For buyers of training data, it turns provenance records into a compliance input. For sellers, it determines what a buyer will disclose about their data.

This guide is based on the text of Regulation (EU) 2024/1689 on EUR-Lex, the Commission's explanatory notice and template, its questions and answers on general-purpose AI and the Digital Omnibus on AI. It is a summary, not legal advice.

## What does Article 53 of the EU AI Act require?

Article 53 of the EU AI Act sets four obligations for providers of general-purpose AI models. Two concern documentation for authorities and downstream users, and two concern training data.

- Article 53(1)(a): keep technical documentation of the model, including its training and testing process, with at least the information in Annex XI
- Article 53(1)(b): give information and documentation to providers who integrate the model into their own AI systems
- Article 53(1)(c): put in place a policy to comply with EU copyright law, including identifying and complying with rights reservations under Article 4(3) of Directive (EU) 2019/790
- Article 53(1)(d): draw up and publish a summary of the content used for training, using the AI Office's template

> "draw up and make publicly available a sufficiently detailed summary about the content used for training of the general-purpose AI model, according to a template provided by the AI Office" (Article 53(1)(d), Regulation (EU) 2024/1689)

Annex XI adds that the technical documentation must describe the data used for training, testing and validation, including its type and provenance, curation methods, number of data points and how it was obtained and selected. That documentation goes to the AI Office on request, while the summary is public.

## Who has to publish a training data summary?

Every provider of a general-purpose AI model placed on the EU market has to publish a training data summary. Article 3(63) defines such a model as one that displays significant generality and can competently perform a wide range of distinct tasks, excluding models used only for research, development or prototyping before they are placed on the market.

- Compute guide: the [Commission's questions and answers](https://digital-strategy.ec.europa.eu/en/faqs/guidelines-obligations-gpai-providers) say models trained with more than 10^23 FLOP that can generate language, text-to-image or text-to-video are indicatively general-purpose, though this is not an absolute rule.
- Open-source models: Article 53(2) exempts qualifying open-source models from points (a) and (b) only. The copyright policy and the training summary still apply, as recital 104 confirms.
- Modified models: a downstream company that modifies a model becomes its provider only in exceptional cases, when the change uses more than one third of the original model's training compute, according to the same questions and answers. Its summary then covers only the data used for the modification.
- Providers outside the EU: Article 54 requires providers established in third countries, including the UK and US, to appoint an authorized representative in the EU before placing a model on the market.

Models trained with more than 10^25 FLOP are presumed to carry systemic risk and face extra obligations under Article 55. Those do not change the training summary itself.

## When do the training data obligations apply?

The training data obligations have applied since August 2, 2025, and the Commission has been able to enforce them with fines since August 2, 2026. Article 113 sets the dates, and Article 111(3) gives older models a longer runway.

| Date | What happened or happens |
| --- | --- |
| August 1, 2024 | The AI Act entered into force |
| July 10, 2025 | The Commission published the General-Purpose AI Code of Practice |
| July 24, 2025 | The Commission published the explanatory notice and training content summary template |
| August 2, 2025 | Article 53 obligations began to apply |
| December 5, 2025 | The notice and template were issued as Commission Communication C(2025) 8311 |
| July 8, 2026 | The Digital Omnibus on AI was adopted, entering into force on July 27, 2026 |
| August 2, 2026 | Commission enforcement powers and fines under Article 101 began to apply |
| August 2, 2027 | Deadline for models placed on the market before August 2, 2025 |

Under Article 101, the Commission may fine a provider up to 3% of its annual total worldwide turnover in the preceding financial year or €15 million, whichever is higher. The explanatory notice confirms that a non-compliant summary can be sanctioned at this level.

## What does the Commission's training data template ask for?

The Commission's training data template asks for three sections: general information about the model and its data, a list of data sources and a description of data processing. The [template and explanatory notice](https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models) cover all stages of training, from pre-training to fine-tuning and alignment.

Section 1 asks for the provider and model identity, the modalities in the training data and their size within broad ranges. Text is reported as less than 1 billion tokens, 1 billion to 10 trillion tokens or more than 10 trillion. Images are reported as less than 1 million, 1 million to 1 billion or more than 1 billion. Audio and video are reported as less than 10,000 hours, 10,000 to 1 million hours or more than 1 million hours. Providers also give the latest collection date and the languages covered.

| Template section | Covers | What must be disclosed |
| --- | --- | --- |
| 2.1 Publicly available datasets | Free, downloadable third-party datasets such as Common Crawl snapshots | Name and link for each large dataset (over 3% of public data for a modality), general description of the rest |
| 2.2.1 Commercially licensed datasets | Data licensed from rights holders or their representatives | Yes or no, and the modalities covered |
| 2.2.2 Other private datasets | Data from intermediaries or private databases without a license from rights holders | Names if publicly known, otherwise a general description |
| 2.3 Crawled and scraped data | Data the provider collected online itself | Crawler names, purpose, behavior, collection period and the top 10% of domains by content size |
| 2.4 User data | Data from users of the provider's services and models | Whether used, and which services it came from |
| 2.5 Synthetic data | Data generated by AI models for training, including distillation | The models used to generate it |
| 2.6 Other sources | Offline sources, self-digitized media, commissioned human labeling, human reinforcement learning data | A narrative description |

Small and medium-sized enterprises, including start-ups, may list the top 5% of scraped domains or 1,000 domains, whichever is lower. Section 3 then asks how the provider respected text and data mining opt-outs, whether it signed the Code of Practice and what it did to remove illegal content from the training data.

## How much detail does the summary need?

The summary needs to be comprehensive in scope but not technically detailed. Recital 107 of the AI Act says it should help parties with legitimate interests, including copyright holders, to exercise their rights, for example by listing the main data collections and explaining other sources in narrative form.

> "this summary should be generally comprehensive in its scope instead of technically detailed" (Recital 107, Regulation (EU) 2024/1689)

The explanatory notice says the template does not require disclosure of the specific works used or the exact mix of data sources. Recital 108 adds that the AI Office will not carry out a work-by-work assessment of copyright compliance. The notice also recommends a voluntary mechanism for rights holders to ask whether content from a specific domain was used, where that domain is not listed.

## How often must the summary be updated?

The summary must be updated when the provider trains its model further on data that changes the summary's content. The explanatory notice sets six-month intervals, or sooner if the new data requires a materially significant update.

The summary must be public no later than the date the model is placed on the EU market. It should sit on the provider's official website and travel with the model across its public distribution channels. The same summary may cover several models or versions if the content is identical.

## What changed in 2026?

In 2026 the training data rules moved from paper to enforcement, while the legislation around them was simplified. The core Article 53 duties did not change.

- Enforcement: from August 2, 2026, the Commission can request information, require corrective measures and impose fines on providers of general-purpose AI models.
- Digital Omnibus on AI: [Regulation (EU) 2026/1744](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng), adopted on July 8, 2026 and published on July 24, 2026, moved the main high-risk obligations to December 2, 2027 for Annex III systems and August 2, 2028 for Annex I systems.
- Content marking: the same regulation gives generative AI systems placed on the market before August 2, 2026 until December 2, 2026 to meet the Article 50(2) marking duty.
- Codes of practice: the regulation removed the requirement for codes of practice to be approved by implementing act. Providers rely on codes the Commission assesses as adequate under Article 56(6).
- Unchanged: the omnibus did not amend Article 53 or the Article 101 fines, so the training summary duty and its deadlines stand as originally set.

The [Code of Practice page](https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai) lists signatories including Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI. It notes that xAI signed only the safety and security chapter, so it must show compliance with the transparency and copyright obligations by other adequate means.

## What does the template mean when you buy training data?

When you buy training data, the template means you need the seller's information in a form you can report. Collect it at diligence, because reconstructing it after training is slow and sometimes impossible.

- The modality and approximate size of the data, in the template's units
- The languages covered and the latest collection date
- Whether the license is with the rights holder or their representative (Section 2.2.1) or with an intermediary that holds no such license (Section 2.2.2)
- For any scraped component, how text and data mining opt-outs were respected
- Any synthetic data and the models that generated it
- Measures taken to remove illegal content

Licensed data carries the lightest public disclosure. The explanatory notice explains that limited detail is required for licensed data because the rights holders are already parties to the license. Data from intermediaries that lack a license from rights holders must be named if publicly known. Our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) and [data provenance guide](https://sellmydatatoai.net/guides/data-provenance-explained) cover how to gather this.

## What does the template mean for data sellers?

For data sellers, the template means a buyer that provides a general-purpose AI model will report your data in its public summary, but usually without naming you. A license signed directly with the rights holder falls under Section 2.2.1, which asks only whether such data was used and in which modalities.

Expect buyers to ask for the information in the list above, and to want warranties that it is accurate. If your data includes material you scraped or bought from elsewhere, expect closer questions, because that data can push a buyer into the more detailed sections. Our guide to [whether you can legally sell company data for AI](https://sellmydatatoai.net/guides/can-you-legally-sell-company-data-for-ai) covers the seller's side.

## How does Sell My Data To AI fit with the EU AI Act?

Sell My Data To AI fits with the EU AI Act as a broker that helps both sides document a deal properly. Sellers sign the license directly with the buyer, and we do not take custody of raw data. Every seller passes our eight-point vetting standard, which covers ownership, third-party content and manifest accuracy.

We are not a law firm, and the provider remains responsible for its own summary. Take independent legal advice on your obligations. To discuss a requirement, [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) or read [how to buy AI training data](https://sellmydatatoai.net/guides/how-to-buy-ai-training-data).

## Frequently asked questions

### Does the EU AI Act training summary apply to UK and US companies?

The EU AI Act training summary applies to any provider placing a general-purpose AI model on the EU market, wherever the provider is based. Article 54 requires providers established outside the EU to appoint an authorized representative in the EU before doing so. A UK or US model provider with EU customers is therefore covered.

### Do open-source model providers have to publish a training summary?

Open-source model providers do have to publish a training summary. Article 53(2) exempts qualifying open-source models from the technical documentation duties in points (a) and (b), but not from the copyright policy or the public summary. Recital 104 explains that releasing weights does not reveal enough about the training data or copyright compliance.

### Will a buyer name my company in its training data summary?

A buyer is unlikely to name your company if it licensed the data directly from you. Section 2.2.1 of the Commission's template asks only whether commercially licensed data was used and in which modalities. Providers may disclose more voluntarily, so agree in the license whether the buyer may name you.

### What happens if a provider's summary is incomplete?

An incomplete summary can be treated as non-compliance with Article 53(1)(d). From August 2, 2026, the AI Office can request corrective measures, and the Commission can fine providers up to 3% of worldwide annual turnover or €15 million, whichever is higher. Providers of models placed on the market before August 2, 2025 may state and justify information gaps.

### Does fine-tuning a model make you responsible for a training summary?

Fine-tuning a model makes you responsible for a training summary only in exceptional cases. The Commission's questions and answers say a downstream modifier becomes a provider when the modification uses more than one third of the original model's training compute. Your summary then covers only the data used for the modification, with a link to the original summary.

## Sources

- [Regulation (EU) 2024/1689 (Artificial Intelligence Act)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng), EUR-Lex, Publications Office of the European Union, accessed November 2026
- [Explanatory notice and template for the public summary of training content for general-purpose AI models](https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models), European Commission, Shaping Europe's digital future, accessed November 2026
- [Guidelines on obligations for general-purpose AI providers: questions and answers](https://digital-strategy.ec.europa.eu/en/faqs/guidelines-obligations-gpai-providers), European Commission, Shaping Europe's digital future, accessed November 2026
- [The General-Purpose AI Code of Practice](https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai), European Commission, Shaping Europe's digital future, accessed November 2026
- [Regulation (EU) 2026/1744 (Digital Omnibus on AI)](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng), EUR-Lex, Publications Office of the European Union, accessed November 2026

---

Book a free consultation: https://sellmydatatoai.net/book-a-consultation