# AI Act Article 53 training-data summary: a template for legal corpora

> Under Article 53(1)(d) of the AI Act, providers of general-purpose AI models must publish a training-content summary using the AI Office template of 24 July 2025; a licensed legal corpus such as Arieta Data is reported in Section 2.2 (private datasets obtained from third parties), while a model trained only on the free public sample, whose records carry their upstream licences, would report it under Section 2.1 (publicly available datasets).

Source: https://data.arietalabs.com/guides/ai-act-article-53-training-data-summary-template/
Published: 2026-11-13 · Updated: 2026-11-14 · Author: Nejc Fošnarič, Founder, Arieta Labs (https://data.arietalabs.com/about/#nejc-fosnaric/)

Article 53(1)(d) of the AI Act, [Regulation (EU) 2024/1689](https://eur-lex.europa.eu/eli/reg/2024/1689/oj), requires every provider of a general-purpose AI model to publish "a sufficiently detailed summary about the content used for training" using a template from the AI Office. The Commission published that template, with an explanatory notice, on [24 July 2025](https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models). A legal corpus licensed from a data company is reported in Section 2.2, "Private non-publicly available datasets obtained from third parties". The same corpus also feeds Section 1.3 (modalities, size, languages) and Section 3.1 (respect of text and data mining reservations). This guide walks through each field with example text for a European legal corpus.

## Who has to publish a training-content summary?

The obligation applies to providers of general-purpose AI (GPAI) models placed on the EU market. Article 53(2) exempts providers of free and open-source models from the documentation duties in points (a) and (b), but not from the summary in point (d) or the copyright policy in point (c), so open-weight models need a summary too (the exemption does not apply to models with systemic risk).

A company that fine-tunes or otherwise modifies a GPAI model can become the provider of the modified model. The explanatory notice (point 28) says that in that case the summary covers only the training content used for the modification, and must name the original model. A legal-AI company fine-tuning a base model on legislation and case law is in exactly this position if the modification makes it a provider.

A data supplier is not a GPAI provider. Arieta Labs does not publish a summary for its buyers' models; it supplies the facts its buyers need to complete theirs.

## What are the key dates?

The obligation and its supervision are phased:

| Date | What happens | Source |
|---|---|---|
| 24 July 2025 | AI Office template and explanatory notice published | [Commission library page](https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models) |
| 2 August 2025 | Article 53 obligations apply to GPAI providers | AI Act, Article 113 |
| 2 August 2026 | AI Office supervision and enforcement for GPAI rules starts | Explanatory notice, point 33 |
| 2 August 2027 | Deadline for summaries of models placed on the market before 2 August 2025 | Explanatory notice, point 33 |
| Every six months | Update the summary when further training on new data requires it, or sooner for a materially significant update | Explanatory notice, point 29 |

The Commission's [FAQ on the template](https://digital-strategy.ec.europa.eu/en/faqs/template-general-purpose-ai-model-providers-summarise-their-training-content) repeats these points.

## How is the template structured?

The template has three sections, each with an optional box for additional information.

| Section | Title | What it asks for |
|---|---|---|
| 1.1 to 1.3 | General information | Provider and model identification; modalities, size ranges, types of content, latest date of data collection, languages |
| 2.1 | Publicly available datasets | Datasets compiled by third parties, free and downloadable; "large" datasets listed by name and link |
| 2.2 | Private non-publicly available datasets obtained from third parties | 2.2.1 datasets commercially licensed by rightsholders or their representatives; 2.2.2 private datasets obtained from other third parties |
| 2.3 | Data crawled and scraped from online sources | Crawlers, collection period, content description, top 10 percent of domain names (top 5 percent or 1,000 domains for SMEs) |
| 2.4 to 2.6 | User data, synthetic data, other sources | Description per source type |
| 3.1 | Respect of reservation of rights from the TDM exception | Measures to honour Article 4(3) reservations of [Directive (EU) 2019/790](https://eur-lex.europa.eu/eli/dir/2019/790/oj), including opt-outs honoured by third parties from which datasets were obtained |
| 3.2 | Removal of illegal content | General description of measures |
| 3.3 | Other information | Optional |

Section 2 asks each dataset to be reported once, in its most relevant category. Filtering a dataset does not turn it into a new dataset.

## Which section does a licensed legal corpus belong in?

A commercially licensed legal corpus belongs in Section 2.2. Which subsection is less obvious, and the answer depends on who holds the rights.

Section 2.2.1 is for commercial licensing agreements concluded "between the provider and the rightsholders or their representatives". Section 2.2.2 covers other private datasets obtained from third parties, which the template describes as including "providers of private databases, or data intermediaries". A corpus of official texts compiled by a company that does not own the rights in those texts, such as Arieta Data, fits 2.2.2 more naturally: the template lists "providers of private databases, or data intermediaries" there, and the explanatory notice (point 19) ties the lighter 2.2.1 disclosure to agreements to which the rightsholders themselves are parties. Arieta licenses its own compilation and enrichment, so some providers may still choose 2.2.1. Take a view with counsel and apply it consistently.

The two subsections differ in disclosure. Under 2.2.1 you tick yes and give the modality. Under 2.2.2 you list the dataset by name if it is publicly known, and otherwise describe its modality, the nature of its content (including personal data) and its languages. Naming the dataset is the simplest way to satisfy either.

The Arieta Data open sample is downloadable from the dataset pages, with each record under its upstream licence; it contains no case law. CJEU case law is in the licensed corpus, and a CJEU case-law excerpt is available under an evaluation licence on request. A model trained only on that public sample would report it under Section 2.1; a copy obtained privately, or the licensed snapshot, is a private dataset under 2.2. The sample will almost never be "large" under the template's test (more than 3 percent of the publicly available datasets used for that modality).

If you scraped EUR-Lex or national portals yourself, or used a web crawl such as Common Crawl ([compared here](/guides/common-crawl-vs-curated-legal-data/)), that data belongs in Section 2.3, with crawler names, the collection period and, where in the top 10 percent, domains such as eur-lex.europa.eu.

## What should each field say for a legal corpus?

The fields below are the ones a legal corpus feeds. The values in the right-hand column are illustrative; use the figures of the snapshot you actually trained on.

| Template field | What to write for a legal corpus |
|---|---|
| 1.3 Modality and size | Text. Size range by tokens across all training data; legal corpora are a component, not the total |
| 1.3 Types of content | The template's own examples include "legal and official documents" |
| 1.3 Latest date of data collection | Month and year of the latest snapshot used, for Arieta Data the [snapshot](/glossary/snapshot/) id in YYYY-MM form |
| 1.3 Linguistic characteristics | The EU official languages and national languages covered |
| 2.2 Dataset name and content | Licensor, dataset name, data type (legislation, court decisions, metadata, citation edges), time span |
| 2.2 Nature of content | Personal data: CJEU judgments name parties; state whether national case law was included |
| 3.1 TDM reservations | How reservations were honoured upstream, and that official texts were sourced from publishers' documented reuse channels |

A short paragraph for Section 2.2.2 might read:

```text
Arieta Data (Arieta Labs Ltd, England and Wales), snapshot 2026-11.
Text: EU legislation and CJEU case law from EUR-Lex and Cellar, reused under
the EUR-Lex legal notice (Commission documents: Decision 2011/833/EU) and the
CURIA copyright notice, and national legislation from official gazettes and
legal databases under their published terms, at article level with metadata,
version history and citation edges. Languages: as listed in the snapshot
datasheet.
Personal data: CJEU judgments as published by the Court; national case law
not included. Raw upstream web pages not included.
```

## What about the copyright policy and TDM opt-outs?

Article 53(1)(c) requires a separate copyright policy that identifies and complies with Article 4(3) reservations of rights, and Section 3.1 of the summary asks you to describe those measures. For licensed data, the template asks how opt-outs were honoured, including by the third parties from which datasets were obtained. Ask your supplier for its per-source record of terms, robots.txt signals and takedowns. For EUR-Lex, reuse rests on an express licence; see [Is EUR-Lex data licensable for AI training?](/guides/eur-lex-data-licence-ai-training/). Why opt-outs have little to bite on for statutes and judgments, and what the GPAI Code of Practice asks for, is covered in [EU copyright, TDM opt-outs and the AI Act for legal data](/guides/ai-act-copyright-tdm-opt-out-legal-data/).

## Key facts

- Article 53(1)(d) of Regulation (EU) 2024/1689 requires every GPAI model provider to publish a training-content summary using the AI Office template of 24 July 2025.
- The obligation applies from 2 August 2025; AI Office enforcement starts on 2 August 2026; models placed on the market before 2 August 2025 have until 2 August 2027.
- A licensed legal corpus is reported in Section 2.2; data you crawled yourself goes in Section 2.3.
- Section 3.1 asks how TDM reservations were honoured, including by the third parties you obtained datasets from.
- Arieta Data's paste-ready text is generated from the snapshot's own statistics; snapshot 2026-11 holds 419,355 documents and approx. 2.05B tokens.

## Frequently asked questions

### Do fine-tuners have to publish a training-content summary?

Only if the modification makes them the provider of a general-purpose AI model. In that case, the explanatory notice says the summary covers only the content used for the modification and must name the original model.

### Which section does a licensed legal dataset go in?

Section 2.2, private datasets obtained from third parties. A corpus compiled by a company that does not own the rights in the texts fits 2.2.2 more naturally; some providers choose 2.2.1. Apply one view consistently.

### Does an open-source model need a summary?

Yes. Article 53(2) exempts free and open-source models from the documentation duties in points (a) and (b), but not from the copyright policy in point (c) or the summary in point (d).

## How Arieta Data supports the summary

Arieta Data publishes a paste-ready description of the current snapshot on the [provenance page](/provenance/#ai-act), generated from the snapshot's own statistics so that figures match the data you received. Each snapshot also ships a [datasheet](/glossary/datasheet/), a per-source LICENCES.md, a MANIFEST.json with SHA-256 checksums, and [Croissant](/glossary/croissant/) metadata following the [MLCommons Croissant specification](https://docs.mlcommons.org/croissant/docs/croissant-spec.html). Datasets are listed on the [datasets page](/datasets/), and per-lane counts for snapshot 2026-11 are on the [coverage page](/coverage/). [Request a sample](/request-access/) and the per-source record: we reply within one business day. See also the glossary entry on the [training-content summary](/glossary/gpai-training-summary/).

*Not legal advice. The template and explanatory notice are the primary sources; take your own legal advice on how to classify each dataset.*