training data summary

advancedspreading

A public document in which a general-purpose AI model developer outlines, as required by EU regulations, on what content the model was trained.

In plain terms

On food packaging, there is a list of ingredients: not the exact recipe, but you can learn from it what the product is made of. The training data summary is the same for AI models: it does not list every single text and image, but shows what sources the training material consisted of.

The EU Artificial Intelligence Regulation (AI Act) mandates that providers of general-purpose AI models, i.e., companies creating the large models behind chatbots and image generators, must fill out a summary using a unified EU template and make it available to anyone.

It is mainly useful for authors, publishers, photographers, and other rights holders to see whether their work could have been included in a model's training data, and if necessary, to assert their rights. It is not a technical description or a complete data list: it is intentionally at a summary level to keep trade secrets protected.

How it works

The developer reviews what the model learned from: public datasets, licensed content, web-scraped pages, user-generated data, and artificially generated (synthetic) data.
Filling out the template issued by the European Commission's AI Office: it provides the model's basic data, the types of training data (text, image, audio, video), and its approximate size.
For content collected from the web, it lists the highest-weighted internet domains and describes how it collected the data and how it took copyright objections (opt-out) into account.
The summary is published on the developer's own website and updated if the model is further trained on new data: at least every six months, or sooner in case of significant changes.
Compliance is monitored by the European Commission through its AI Office, which can impose fines for non-compliance: the upper limit is 3 percent of the previous year's global market turnover or 15 million euros, whichever is higher.

An everyday example

As a journalist or photographer, you may wonder whether a popular chatbot learned from your work. You visit the developer's website, look up the model's training data summary, and check whether the portal where you publish is listed among the major web sources.

Who is affected

lawyer, citizenAs an author or rights holder, this shows you where to look into whether your works were used, and it can provide a basis for enforcing your rights. → Law

developerIf you create or significantly further train a general-purpose model, you must prepare and publish the summary. → Code

entrepreneur, executiveWhen selecting an AI model for your company, the summary allows you to assess how transparent and legally compliant the supplier's data background is.

everyoneAs a user, you can see whether the provider included user data in the training.

A common misconception

MythFrom the summary, one can look up exactly whether my article or photo was in the training data.

In realityThe summary mentions only source categories, larger datasets, and the most important domains, not individual works. It provides a pointer but not evidence on its own.

Limits and risks

The summary is the developer's own statement: its content is not pre-checked by an independent organization, though authorities may review it afterward.
The level of detail is limited because the rule also protects business secrets, so smaller sources typically remain hidden.
It applies only to general-purpose models available on the EU market, and models launched before 2 August 2025 were granted a deadline extension until 2 August 2027.
The mere fact that a source is included does not imply that a violation occurred; that can be decided by a separate legal procedure.

Similar, but not the same

model card – The model card describes the model's capabilities, limitations, and recommended use, typically voluntarily, while the training data summary is mandatory and concerns only the origin of the training material.
dataset card – A data sheet documents a single dataset in detail as a research practice, while a summary presents a model's entire training material at an overview level as a legal requirement.
technical documentation – The technical documentation is detailed, non-public, and prepared for authorities (companies embedding the model receive separate information), whereas the summary is shorter and readable by anyone.

Related concepts

Broader

transparency, EU MI-rendelet, AI regulation

Narrower

On the edge

training data, general-purpose AI model, model card, adateredet, synthetic data

A brief history

The obligation was introduced by the EU's Artificial Intelligence Act adopted in 2024, in Article 53(1)(d). The European Commission published the mandatory template in July 2025, and the rule applies to new general-purpose models from August 2, 2025.

Where things stand

By the end of 2026, the obligation already applies to new general-purpose models, and the European Commission has been able to impose fines through the AI Office since August 2, 2026. Models previously placed on the market must comply by August 2, 2027.

Status: 2026 November

Also known as

training data summary · summary of learning materials · public summary of training data · summary of training data · training data summary · training content summary · public summary of training content · AI Act training data template

Sources

Wikipedia: Artificial Intelligence Act

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.