Copyright and AI law

EU AI Act and training data: what the transparency rules require

An angled view of the ABC cinema and traffic light trails beneath a purple evening sky.
ABC cinema at dusk.Photograph by Robert Harper. Robert Harper Photograph Library, Archive.London. From the archive, shown for illustration.

Since August 2025, anyone placing a general-purpose AI model on the EU market has had to publish a summary of what it was trained on. Since August 2026, the Commission can fine those who do not.

The EU AI Act does not tell AI developers what they may train on. Copyright law does that. What the Act adds is a duty to have a policy for obeying copyright law and a duty to tell the public, in summary, where the training data came from. Both have applied to providers of general-purpose AI models since 2 August 2025, and the European Commission has been able to enforce them since 2 August 2026.

This article is general information, not legal advice. Checked on 18 November 2026.

What does Article 53 require?

Article 53 of Regulation (EU) 2024/1689 sets obligations for providers of general-purpose AI models, the large models that can be adapted to many tasks. Two sub-paragraphs concern training data.

Article 53(1)(c) requires a policy to comply with EU copyright law, and in particular to identify and respect rights reservations made under Article 4(3) of the 2019 Copyright in the Digital Single Market Directive. That is the provision that lets rights holders opt out of commercial text and data mining by reserving their rights "in an appropriate manner, such as machine-readable means". How that works, and how it differs from British and American law, is set out in our comparison of text and data mining exceptions.

Article 53(1)(d) requires the provider to draw up and publish a "sufficiently detailed summary" of the content used for training, following a template supplied by the AI Office.

Recital 106 adds that providers placing a model on the EU market should respect these copyright rules regardless of the jurisdiction in which training took place. A recital is an aid to interpretation, not an operative provision, and how far that statement carries is not yet tested in court.

What goes in the training content summary?

The Commission published the template and an explanatory notice on 24 July 2025. According to the Commission's own FAQ, the summary has three parts.

Section What the provider must give
General information Provider and model identity; the modalities of training content, such as text, image, video and audio; training data size per modality within broad ranges
List of data sources Large publicly available datasets, listed; private datasets, described; scraped online data with crawler names, collection period and the top 10% of domains (top 5% or 1,000 domains, whichever is lower, for small and medium-sized providers); whether user data was used; synthetic data
Data processing aspects Information relevant to rights holders and to compliance with the lawful conditions for text and data mining; removal of illegal content

The summary must be on the provider's website, clearly visible, and alongside the model on its public distribution channels, no later than when the model is placed on the EU market. It must be updated every six months, or sooner if further training makes a material difference.

The Commission describes the template as a "common minimal baseline" and says it balances transparency against trade secrets by asking for different levels of detail for different sources.

Does a provider have to list every image it trained on?

No, and this is the most common misreading of the Act. The template works at the level of datasets and domains. A photographer cannot look up a single picture in a provider's summary. What a rights holder can learn is whether a named public dataset was used, which crawlers collected data and when, and whether their website falls among the most heavily scraped domains.

For an image library that is still useful. If a summary names a public dataset known to contain the library's pictures, that is a factual starting point for a conversation or a claim. For a data-procurement team on the other side of the table, the template creates a paper trail: privately licensed datasets have to be described, so the terms on which they were acquired become something the provider must be able to stand behind.

When did the rules start, and who enforces them?

Date Event
10 July 2025 General-Purpose AI Code of Practice published
24 July 2025 Training content template published
2 August 2025 Article 53 obligations begin to apply
2 August 2026 Commission enforcement powers, including fines, begin
2 August 2027 Deadline for models placed on the market before 2 August 2025

Enforcement sits with the Commission, acting through the AI Office. Fines for general-purpose model providers can reach 3% of annual worldwide turnover or EUR 15 million, whichever is higher.

What has happened since August 2026. On 14 October 2026 the Commission's executive vice-president Henna Virkkunen announced that the AI Office had sent its first formal requests for information. According to Reed Smith's account, they went to providers of the most advanced models and concerned security practices, external evaluation and post-market monitoring. Copyright and training summaries were not the reported subject. We found no report of a fine or a formal copyright investigation under Article 53 as of the date of this article.

How providers have responded. One comparison of published summaries found that Google, Meta, Microsoft and OpenAI had completed the template, while Anthropic, Mistral and xAI had published narrative statements instead. That comes from a single secondary source, published on 7 August 2026, which we have not checked against the summaries themselves. The same source notes that many of the models concerned predate 2 August 2025, so their deadline falls in 2027, and it does not say any provider is in breach. It should be treated as indicative, not as a compliance finding.

The Code of Practice is a voluntary way to demonstrate compliance with Article 53. The Commission endorsed it as adequate on 1 August 2025. Its copyright chapter commits signatories to keep a copyright policy, to crawl only lawfully accessible content, not to circumvent paywalls, to honour robots.txt and other recognised machine-readable rights reservations, to take steps against infringing outputs, and to run a contact point for complaints.

Not everyone signed. Meta declined, and xAI signed only the safety and security chapter.

The unsettled part is which opt-out signals count. The Commission consulted between December 2025 and 23 January 2026 on the machine-readable protocols that providers must respect. Seven were assessed, and the picture-industry body CEPIC reported strong disagreement among rights holders. We could not verify a final list. robots.txt is the only protocol the Code itself names, and it operates per website, not per image. German courts have also had something to say about what "machine-readable" means, which we cover in our article on Kneschke v LAION and machine-readable opt-outs.

Did the Digital Omnibus delay any of this?

The Digital Omnibus on AI changed several AI Act deadlines. On the secondary summaries available, it pushed obligations for stand-alone high-risk systems to 2 December 2027 and for high-risk systems embedded in products to 2 August 2028, and gave systems already on the market until 2 December 2026 to meet certain content-marking duties. Political agreement was reported on 7 May 2026 and publication in the Official Journal on 24 July 2026 as Regulation (EU) 2026/1744; we have not checked those details against EUR-Lex.

None of the reports we reviewed describes any delay to Article 53. The copyright policy and training summary duties continued on their original timetable.

What is decided and what is still open?

Decided or in force: the Article 53 duties; the template; the Commission's power to enforce; the 2027 deadline for older models.

Pending or uncertain:

  • The list of opt-out protocols that providers must honour.
  • Whether, and against whom, the AI Office will use its powers on copyright.
  • The Court of Justice's first ruling on generative AI. In Like Company v Google Ireland (C-250/25), Advocate General Szpunar advised on 3 November 2026 that the questions about training should be declared inadmissible as hypothetical, while adding for completeness that training a language model is as a rule permitted under the Article 4 exception. An Advocate General's opinion does not bind the Court, and judgment will come later.
  • A review of the 2019 Directive itself, which Article 30 of the Directive requires the Commission to carry out no sooner than 7 June 2026. We could not confirm a publication date.

What this means for people who license photographs

The Act rewards clean inputs. A provider that licenses a described, documented image collection from its owner can state exactly that in its summary. A provider relying on scraped data has to name its crawlers and its most-scraped domains, and has to show how it respected reservations.

For libraries, the general lesson is that reservations machines can read count for more than those they cannot. Site-level signals protect a website; they do not travel with a file once it is copied elsewhere. The IPTC "Data Mining" property is a signal designed to sit inside an image file, and our free AI training opt-out metadata tool will write it into a JPEG without uploading the file anywhere. It disappears if metadata is stripped, which is a limit worth knowing.

One archive, one counterparty

The Robert Harper Photograph Library holds 95,487 unique photographs by the archive's inventory count, 52,684 of them analogue, with one photographer behind them and one owner and rights holder, 825 Ventures Ltd. The catalogue runs to 330 described collections, from studio still life to the corporate and technology collections.

For a provider that must describe its private datasets, a single-source collection with a documented origin is straightforward to describe. Licences are negotiated directly with the rights holder and are subject to rights review. The AI training data licensing page explains what is available and how to ask.

Frequently asked questions

What is the EU AI Act training data summary?

It is a public document that every provider of a general-purpose AI model placed on the EU market must publish under Article 53(1)(d) of the AI Act, using a template from the AI Office. It describes the kinds and sources of training content: public datasets, private datasets, scraped web data, user data and synthetic data.

Does the EU AI Act require AI companies to list every work they trained on?

No. The template asks for information at the level of datasets and sources. For scraped web data, providers list crawler names, the collection period and the top 10% of domains; smaller companies list the top 5% or 1,000 domains, whichever is lower. Individual photographs or articles are not itemised.

Does the AI Act apply to models trained outside the EU?

The copyright policy duty is tied to placing a model on the EU market. Recital 106 of the Act says providers should comply with the Union's copyright rules on rights reservations regardless of where training took place. A recital guides interpretation and is not itself an operative rule, so how far this reaches is likely to be tested.

Has anyone been fined under Article 53 yet?

We found no report of a fine or formal copyright investigation as of 18 November 2026. The AI Office's first formal requests for information, announced on 14 October 2026, concerned security practices, external evaluation and post-market monitoring at providers of the most advanced models, according to Reed Smith.

On the summaries we have seen, no. The Digital Omnibus on AI postponed obligations for high-risk systems and gave a grace period for some content-marking duties. The Article 53 obligations for general-purpose model providers, including the copyright policy and training summary, were not reported as delayed.

Sources

This article is general information, not legal advice. It reflects the position as we understood it on the date shown and may be out of date. Published by Archive.London. Corrections are welcome under our editorial policy.

Licensing the archive for AI and data projects

95,487 photographs from one photographer and one rights holder. Read the dataset card, search the 503-folder inventory, discuss a sample set under NDA, or arrange a first look as scanning proceeds.

See AI and data licensingWrite to the archive