EU AI Act Article 53: The Training Data Summary and Copyright Policy

Article 53 of the EU AI Act requires every provider of a general-purpose AI model to keep a copyright policy and to publish a summary of the content used to train the model, so a company that licenses data to such a provider meets the law through the buyer's disclosures, not through its own.

The rules sit in Regulation (EU) 2024/1689 (CELEX 32024R1689), in force since 2024-08-01, with Article 53 applying from 2025-08-02. This page covers the duties, the summary template, the Code of Practice, the dates, a published example and the licensor's view as of 2026-11-09, for anyone weighing AI training data licensing in the EU.

Every rule below cites the Official Journal text, the Commission's template or a provider's own summary; the method for sourcing and dating every figure is public.

What Article 53 Requires of a General-Purpose AI Model Provider

Article 53(1) sets four duties for every provider of a general-purpose AI model: technical documentation, information for downstream providers, a copyright policy and a public summary of training content. The table lists them in article order, as of 2026-11-09.

Points (c) and (d) are the two duties that reach licensed data. Article 53(4) lets providers rely on a code of practice to show compliance until a harmonised standard exists.

Which Providers Carry the Duties: Compute Thresholds and Open-Source Models

The Commission's guidelines treat a model as general-purpose indicatively from 10^23 FLOP of training compute, and presume systemic risk from 10^25 FLOP, a threshold the Commission describes as under review (factpage checked 2026-11-09). Article 53(2) exempts free and open-source models with public weights only from points (a) and (b), never from (c) and (d), and not at all where the model has systemic risk.

When the Duties Apply: August 2, 2025 to August 2, 2027

Article 53 has applied since August 2, 2025, but the Commission can fine providers only from August 2, 2026, and models placed on the market before August 2, 2025 have until August 2, 2027. The table lists seven dates in order, as of 2026-11-09.

Regulation (EU) 2026/1744, the Digital Omnibus on AI, moved the high-risk dates but did not amend Article 53; only the Article 97 references to Article 53(5) and (6) were touched.

The Public Summary of Training Content Under Template C(2025) 8311

The summary under Article 53(1)(d) must follow a template from the AI Office, and the current version is Commission Communication C(2025) 8311 final of December 5, 2025, first published on July 24, 2025. The AI Office is the Commission service that issued it and supervises general-purpose AI models.

The law asks for a "sufficiently detailed summary", never a list of every dataset. The template sorts the main data sources into licensed data, other private datasets and scraped web content, and the library page for it was last updated on 2026-03-26.

Licensed Datasets: Yes or No Plus Modality (Section 2.2.1)

For data obtained under commercial licenses with rightsholders or their representatives, the provider states only Yes or No and the modality: text, image, video, audio or other. Paragraph 19 of the explanatory notice gives the reason: the rightsholders are parties to those licenses, so the disclosure stays limited.

For the licensor this means it can expect not to be named in the public summary, and a data manifest can mirror the template's modality and size bands.

Datasets Bought Through Intermediaries (Section 2.2.2)

Private datasets obtained from other third parties, such as data intermediaries, are disclosed as Yes or No plus modality, and they are listed only if they are already publicly known, otherwise described in general terms. A licensor selling through a broker or marketplace is therefore not named unless the dataset is already public.

Scraped Domains and Six-Month Updates

For web-scraped content, providers that are not SMEs list the top 10% of domain names by size of content scraped. Two further rules apply.

• SMEs list a smaller set, and the two texts differ: the explanatory notice reads "top 5%, or the top 1000 domains", while the annexed template adds "whichever is lower".
• The summary is updated at six-month intervals, or earlier if additional training is materially significant (paragraph 29).

A Published Summary Read Section by Section: Grok 4.7

The public summary for Grok 4.7, version 1 dated September 21, 2026, shows how a provider fills the template, and every entry below is SpaceXAI's own statement. The table maps eight entries to the template, as of 2026-11-09.

The summary says Yes to commercial licenses for text and image without naming any licensor, which is the Section 2.2.1 rule at work. The summary of Grok 4.5 (2026-07-08) was not read for this page, and no other provider's summary is described here; the sourcing table on how frontier AI labs as data buyers disclose their sources keeps the lab comparison.

The Copyright Policy Under Article 53(1)(c) and Rights Reservations

Article 53(1)(c) requires a policy to comply with Union copyright law, in particular to identify and comply with rights reservations under Article 4(3) of the DSM Directive (Directive (EU) 2019/790), including through state-of-the-art technologies. Three points define the duty.

• Article 4(3) allows a reservation for online content in an appropriate manner, such as machine-readable means.
• Article 7(1) protects only Articles 3, 5 and 6 from contract override, so a license can set different terms than the Article 4 exception.
• Recital 106 states that the policy duty applies regardless of the jurisdiction in which the training acts took place.

How the reservation works is covered on the page on the EU text and data mining exception and its opt-out.

The Code of Practice Copyright Chapter and Its Signatories

The General-Purpose AI Code of Practice, published on July 10, 2025, is a voluntary code whose Copyright chapter sets out how signatories meet Article 53(1)(c). Measure 1.2 commits signatories to respect paywall and subscription restrictions and to exclude sites that EU or EEA courts or authorities recognize as persistently infringing on a commercial scale. Measure 1.3 commits their crawlers to read and follow robots.txt under IETF RFC 9309 and other machine-readable Article 4(3) protocols.

The table shows the signatory status on the Commission page, last updated 2026-10-07 and checked 2026-11-09; the 22 listed names are in the source's alphabetical order.

Signing is voluntary; under Article 53(4) a provider that does not adhere must demonstrate alternative adequate means of compliance for assessment by the Commission. The robots.txt commitment is where crawler names matter, and the page on AI crawlers lists which bots collect training data and how robots.txt reaches them.

Enforcement: Article 101 Fines From August 2, 2026

The Commission can fine a general-purpose AI model provider up to 3% of annual total worldwide turnover or EUR 15,000,000, whichever is higher, under Article 101(1), and that power applies from August 2, 2026. Article 113(b) excludes Article 101 from the 2025-08-02 application date, and paragraph 33 of the template places AI Office supervision and enforcement from 2026-08-02. This page states no fine or proceeding under Article 101.

What Article 53 Means for a Data Licensor: Benefits, Risks, Downsides and Alternatives

Article 53 binds the buyer, not the seller, so its effect on a company licensing data is indirect, and licensing to a provider outside its reach, or not licensing at all, remain valid choices.

Benefits

• The licensor is not named in the public summary, only Yes or No plus modality (Section 2.2.1).
• Datasets sold through an intermediary stay unlisted unless they are already public (Section 2.2.2).
• The buyer must keep a copyright policy that honors Article 4(3) reservations.

Risks

• A modality-level Yes still signals that licensed data of that type exists, as the Grok 4.7 entry shows.
• The buyer may ask for terms on rights and reservations because its own policy depends on them; such terms are covered in the guide to AI data licensing agreements.
• Models placed on the market before 2025-08-02 have until 2027-08-02, so older models may not yet carry a summary.

Downsides

• Summaries update only at six-month intervals, so a licensor sees changes late.
• The Code is voluntary, so a signature does not bind every buyer.
• The template gives no list of licensors that a seller could check.

Alternatives

• License only content already free of third-party rights.
• License to a provider that does not place a model on the EU market.
• Wait for the buyer's next summary before agreeing terms.
• Do not license at all.

The choice between these options, including keeping the data, is set out on the page on whether to sell or license your data to AI companies at all.

Beyond Article 53: Copyright, Privacy and Disclosure Rules on Other Pages

Article 53 is one EU duty of the buyer, and the rights in the content, the personal data in it and other disclosure laws each have their own page. The parent page covers EU law for AI training data as a whole.

• AI training data law by jurisdiction for rules outside the EU.
• GDPR and AI training data for personal data in a licensed set.
• Copyright and AI training across jurisdictions for fair use and TDM rules compared.
• How AI companies get training data for the source categories a summary reports.
• California AB 2013 training data transparency for the California disclosure statute.
• X.AI v. Bonta for the challenge to AB 2013.
• Training data transparency rules for disclosure rules and initiatives beyond the EU.
• xAI (SpaceXAI) data buyer profile for the provider of the example summary.
• Alternatives to selling data to AI labs, including not selling.

How This Page Is Sourced and Dated

Every rule here cites the Official Journal text, the Commission's template or Code, or a provider's own published summary, with the date it was checked. The rules are set out in how we source, date and check every figure, and this page is general information about the EU AI Act, not legal advice for any specific deal.

Frequently Asked Questions

Are open-source AI models exempt from Article 53?

No, Article 53(2) exempts free and open-source models only from the documentation duties in points (a) and (b). They still need a copyright policy and a training summary.

How often must a training data summary be updated?

At six-month intervals, or earlier if additional training is materially significant. The rule is in paragraph 29 of C(2025) 8311.

Is the GPAI Code of Practice binding law?

No, it is voluntary. Article 53(4) lets providers rely on it to show compliance until a harmonised standard is published.

How does California's AB 2013 differ from the EU training data summary?

AB 2013 (Cal. Civ. Code 3110-3111) is a separate California statute on training-data disclosure by generative AI developers. A court filing states it is in force from 2026-01-01, and a challenge by X.AI LLC is pending on appeal as of 2026-11-09.