Training data and copyright under the EU AI Act: the Article 53 policy obligation
The EU AI Act ties generative AI training data to copyright law. This guide explains the Article 53 copyright policy obligation, the TDM opt-out rule, and how builders and buyers should respond.
Generative AI in Europe does not get a free pass on copyright. Since 2 August 2025, providers of general-purpose AI models must run a working policy for honouring rights reservations — the machine-readable text and data mining opt out EU copyright law has recognised since 2021 — and publish a summary of what they trained on. This article covers the Article 53 copyright policy requirement, its overlap with the EU copyright directive on AI training, and how to write the policy and the questions you put to a vendor.
What the Article 53 copyright policy requirement actually says
Article 53 of Regulation (EU) 2024/1689 sets the obligations for providers of general-purpose AI models. Paragraph 1 lists four duties, two of which concern copyright:
- Points (a) and (b) — technical documentation on training and evaluation (Annex XI) for the AI Office and national authorities, and information for downstream providers who integrate the model (Annex XII).
- Point (c) — put in place a policy to comply with Union law on copyright and related rights, and in particular to identify and comply with, including through state-of-the-art technologies, a reservation of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790.
- Point (d) — draw up and make publicly available a sufficiently detailed summary of the content used for training, on the template issued by the AI Office.
The wording of (c) matters: the Act does not ask for a statement of intent — it asks that you identify rights reservations and comply with them using current technical means, not a manual skim. Two details catch mid-market teams out. The open-source relief in Article 53(2) exempts only points (a) and (b), so an open-weights model still owes the copyright policy and summary; and compliance may be shown through a code of practice under Article 56. Enforcement runs through Article 101, with fines up to 3% of worldwide annual turnover or €15 million from 2 August 2026.
Where the EU copyright directive and AI training overlap
The AI Act does not create a copyright exception. It points back to one. Directive (EU) 2019/790 on copyright in the Digital Single Market made text and data mining lawful in two layers:
- Article 3 covers mining for scientific research by research organisations and cultural heritage institutions. There is no opt-out: a rightholder cannot reserve rights out of it.
- Article 4 covers mining for any other purpose, including commercial model training, of lawfully accessible works. Paragraph 3 makes the exception conditional: it applies only where the use has not been expressly reserved by the rightholder in an appropriate manner, “such as machine-readable means in the case of content made publicly available online”.
That sentence is the hinge of the regime — what Article 53(1)(c) asks providers to identify and honour. “Lawfully accessible” carries weight too: content behind a paywall, login or technical control may fall outside the exception whatever a robots file says.
Because the directive was transposed nationally, the same reservation appears across German, French, Dutch, Spanish and Italian copyright law. The overlap is practical: the AI Act regulates the provider’s behaviour, the copyright directive supplies the rule it must respect, and failure under either shows up in the same place — the training pipeline.
Can AI train on copyrighted data in EU law?
Yes — but only down three routes, and only if the evidence for your route survives an audit.
- Licence. A negotiated or stock licence covering text and data mining expressly, not merely reading rights — the cleanest route, the one buyers ask about, the only one reliably carrying indemnity.
- The Article 4 exception, with the reservation honoured. Lawfully accessible works whose rightholder has not reserved rights machine-readably may be mined — absence of a reservation is permission by default, but it must be recorded, because “we checked and found nothing” is only credible if someone can show the check.
- Public-domain and openly licensed material. Out of copyright, or licensed to permit the use — simplest, and usually a minority of any real web corpus.
Two failure modes account for most exposure: corpus drift — a crawl taken under one set of terms reused years later, after the site added a reservation — and downstream reuse, where a fine-tuned model inherits whatever its base model’s documentation omits. A retailer training a product-description model on competitor catalogues is doing commercial text and data mining of third-party works, lawful only down route 1, 2 or 3, and answers for it as the party that named the model.
The training-content summary under Article 53(1)(d)
The second duty is public-facing. A provider must publish a sufficiently detailed summary of the content used for training, following the template the AI Office has issued. The template is deliberately coarse — it is a transparency instrument, not a dataset manifest — and it asks for three broad blocks:
- Who and when: model name, provider, training version or date, contact point.
- What: content categories — licensed corpora, user-provided data, public extracts, and for crawled web data the most significant source domains by volume, plus personal-data filtering applied.
- How copyright was handled: which rights reservations were found, how detected, how respected.
Three practical points: the summary must be published, so vague phrasing is read as an admission; it need not be exhaustive, but a category missing from the summary while present in the pipeline is a mismatch regulators can test with one question; and it is the artefact a buyer reads before the contract, answering in public most corpus questions procurement would otherwise ask under NDA.
An AI provider copyright policy template in eight clauses
Policies fail when they read as statements of values. The one below reads as a procedure, which is what Article 53(1)(c) tests — adjust the names, keep the shape.
- Scope and ownership. Applies to every activity acquiring content for training, fine-tuning or retrieval — staff uploads, vendor corpora, partner feeds. Head of Legal owns it, the ML lead executes, the DPO signs off on personal data.
- Lawful basis per corpus. Every dataset is tagged at ingestion as licensed, Article 4 exception applied or public domain / open licence; untagged data may not enter a training run.
- Reservation detection. Before ingestion and at each refresh, sources are checked for express reservations using machine-readable means — robots directives, licence metadata — on a defined cycle.
- Detection result recorded. Each source carries a dated record of the signal found or not; a check that cannot be evidenced is treated as a reservation and the source excluded.
- Technical measures. Filters, provenance tagging and per-source ledger entries implement “state-of-the-art technologies”; the ledger feeds the Article 53(1)(d) summary.
- Escalation and withdrawal. A reservation, complaint or takedown triggers removal from the active corpus within five working days, logging of affected runs, and a refresh assessment.
- Publication and review. Policy and summary are published, reviewed at least annually, and re-reviewed whenever a new model version ships.
- Flow-down. Suppliers, data partners and downstream customers inherit the same constraints contractually, with warranties matching the risk.
Teams that run this shape find clauses 2 and 4 do most of the work: the first stops informal datasets entering the pipeline, the second ends arguments about what had actually been checked.
Text and data mining opt-out and the AI Act: what downstream users must not do
The Article 53 duties sit on providers of general-purpose AI models. The underlying prohibition does not. Copyright law — the transposed Article 4 of the directive — applies to everyone who reproduces or extracts works, including your data team, your agency, and whoever assembled the dataset you bought last quarter.
In practice, downstream circumvention looks ordinary rather than deliberate:
- Configuring a crawler to ignore exclusion signals — disabling robots directives, rotating past a block, or pulling through an endpoint the owner did not offer — is the clearest case of not respecting a machine-readable reservation.
- Assembling a “pre-cleaned” dataset: a scraped bundle with reservations stripped out moves the breach into your pipeline with a receipt attached.
- Prompting for substantial reproduction of reserved works is mining by another name.
- Treating silence as consent indefinitely: a site that reserved nothing in 2023 may reserve today; dated re-checks are the difference between a defensible position and corpus drift.
None of this is excused by a vendor having supplied the base model, and no contract fully transfers it: indemnities shift cost, they do not make the use lawful. What a contract can do is make the vendor warrant what Article 53(1)(c) requires of them, so an upstream failure becomes a breach you can claim on.
Questions to ask your AI vendor about model training data and copyright
Mid-market buyers rarely get the training corpus — but they get answers, and these seven questions separate a documented position from a marketing line. Take them into the security and legal review together.
| Question to ask your AI vendor | Why it matters | Article or directive hook |
|---|---|---|
| Is this a general-purpose AI model, and are you its provider? | Determines who carries the Article 53 duties, and whether your fine-tuning pushes you onto the provider side | AI Act Art. 3(63), Art. 53(1) |
| Where is your copyright policy published? | The duty is to put a policy in place, not assert one; no policy means the starting line has not been crossed | AI Act Art. 53(1)(c) |
| Where is the training-content summary for this version? | A corpus check you can run before the call, no NDA needed | AI Act Art. 53(1)(d) |
| How do you detect machine-readable reservations, and how often do you re-check? | Tests whether “state-of-the-art technologies” means anything; exposes corpus drift | AI Act Art. 53(1)(c); DSM Art. 4(3) |
| What share of training content is licensed rather than exception-mined? | The licensed share is the indemnifiable share; the rest rests on honoured reservations | DSM Art. 4(1) and 4(3) |
| Relying on an approved code of practice or harmonised standard? | Distinguishes third-party assessed compliance from self-declaration | AI Act Art. 53(4), Art. 56 |
| What IP indemnity applies to outputs, and what voids it? | Indemnity is commercial, not statutory; caps and exclusions decide who pays | Contract; no AI Act hook |
Record the answers in your AI register next to the model entry — a vendor answering four of seven in writing tells you something useful about the other three.
How to check your own position in three minutes
The free exposure scan asks twelve questions about your AI usage and governance and returns an exposure score, your ranked exposures with the exact article that fired, and your first three moves mapped to the dates that apply. Confirm your email to unlock the full report.
Frequently asked questions
Does the EU AI Act ban training on copyrighted data?
No — the Act creates no copyright exception and no ban. It requires GPAI providers to run a policy identifying and honouring rights reservations (Article 53(1)(c)) and to publish a training-content summary (Article 53(1)(d)). Lawfulness comes from copyright law: a licence, the DSM Directive Article 4 exception with reservations honoured, or open-licensed material.
What counts as a machine-readable opt-out?
A rights reservation expressed so automated tooling can detect it — robots directives, licence metadata, tags or dataset terms expressly excluding text and data mining. The Act’s phrase “through state-of-the-art technologies” means you must actively detect reservations, not avoid noticing them; document what you check and when — an unevidenced check is not a defence.
Do open-source models escape the copyright policy?
No. The Article 53(2) relief for free open-source models exempts only the documentation and downstream information duties — points (a) and (b). The copyright policy and training-content summary remain owed by every GPAI provider, open-weights or not.
We only fine-tune a third-party model — does this touch us?
Directly, if you assemble training data: copyright law binds your pipeline regardless of who built the base model. Via the Act, if fine-tuning makes you a provider under Article 25. As a minimum, contract for the provider’s policy and summary, and keep your own corpora tagged under routes 1 to 3.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.