This article is about the legal rules for collecting data to train an AI model in the EU. It covers the copyright exception for text and data mining (TDM, the automated analysis of texts and images), the opt-out that rightsholders can declare, the duties of general-purpose AI model providers under the AI Act, the GDPR rules for personal data in training sets and the latest court decisions in Germany and at EU level. Everything is based on legislation, regulator documents and court press releases.
Checked on 28 October 2026. Rules and sources are current as of this date.
Three layers of risk in one dataset
A training set can raise three separate legal questions at once. Each has its own rules, and clearing one does not clear the others.
Copyright
Collecting works and copying them into a dataset is reproduction. In the EU it is lawful only if a licence or an exception covers it. The main exception is the one for text and data mining.
Personal data
Names, photos, posts and voices are personal data. The GDPR (General Data Protection Regulation) applies to collecting them, even from public pages, and to a model that may still contain them.
AI Act
A provider of a general-purpose AI model must keep a copyright policy and publish a summary of its training content. The duty sits with the provider, not with every company that uses a model.
Contracts and site terms
Website terms, API licences and data supply contracts can forbid scraping or AI use even where the law would allow it. A breach is a contract claim.
| What you do | Which rules work | What to check |
|---|---|---|
| Scrape public web pages | TDM exception, GDPR | Opt-out signals, personal data in the pages, site terms |
| Buy or license a dataset | Contract, GDPR | Chain of rights, warranties, where the data came from |
| Use your own users’ content | Contract, GDPR, copyright | Terms of service, user licence, legal basis, notice to users |
This article covers EU law. US courts have in some decisions treated training as fair use (the Bartz and Kadrey cases), which is a different test, and the Munich court distinguished those decisions in the Suno case below.
The TDM exception and the opt-out
The DSM Directive (Directive 2019/790 on copyright in the Digital Single Market) has two exceptions for text and data mining. Member states put them into national law, for example sections 60d and 44b of the German Copyright Act.
Article 3, research
Research organisations and cultural heritage institutions may copy and mine works they have lawful access to for scientific research. Rightsholders cannot opt out of this exception. A commercial company does not qualify on its own.
Article 4, everyone
Anyone with lawful access may copy works and extract information from them, including a commercial AI developer. The condition is that the rightsholder has not reserved the use in an appropriate manner, such as machine-readable means for content published online (Article 4(3)).
Copies are temporary
Copies may be kept only as long as necessary for the mining (Article 4(2)). A permanent library of scraped works is a different thing from a working copy.
What counts as an opt-out
The Directive speaks of machine-readable means. In practice this points to robots.txt (a file that tells crawlers which pages to skip), metadata and other recognised protocols. Whether a sentence in the website terms is enough is disputed, as the Hamburg case below shows.
The exception concerns the copying done while data is collected and analysed. It does not cover what the trained model later outputs. The Munich court said so in two separate cases, described next.
What the courts have decided
There is no final answer at EU level yet. These decisions guide practice today, and none of the German ones is final.
LAION, Hamburg appeal court, 10 December 2025
A photographer sued LAION, a non-profit that built a dataset of image links and descriptions. The court held that downloading the photo to compare it with its description was covered by the TDM exception and also by the research exception. The reservation on the photo agency’s website was not machine-readable, so it did not block the copy. An appeal to the Federal Court of Justice (BGH) is allowed.
GEMA v OpenAI, Munich Regional Court, 11 November 2025
Lyrics of nine songwriters could be reproduced from the GPT-4 and GPT-4o models. The court held that storing them in the model is reproduction. The TDM exception covers preparatory copies made to assemble training data, but not memorisation that lets the model reproduce works. OpenAI and not the user was responsible for the outputs. The judgment is not final.
GEMA v Suno, Munich Regional Court, 31 July 2026
The court largely granted GEMA’s claims for six works in an AI music generator. It found the works memorised, held that the TDM exception did not cover the memorisation in models stored in Germany, and held Suno responsible for the outputs. It also found no fair use for the training done in the US. The judgment is not final.
Like Company v Google, Court of Justice (C-250/25)
The first EU Court of Justice case on generative AI and copyright. A Hungarian publisher asks whether a chatbot’s answers communicate its press publications to the public, whether training is reproduction and whether the TDM exception covers either. The hearing took place on 10 March 2026. We found no judgment as of 28 October 2026.
On these rulings, collecting data for training may be covered by the exception, while a model that memorises works and reproduces them in outputs is not. The risk moves from the collection to the model’s behaviour. Who owns generated output is a separate question, see our article on AI-generated content and copyright.
AI Act duties for model providers
The AI Act (Regulation 2024/1689) adds two duties for providers of general-purpose AI models, meaning models that can perform many different tasks. They have applied since 2 August 2025.
Copyright policy
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 DSM Directive, using state-of-the-art technology. Recital 106 says this applies wherever the training took place.
Training content summary
Article 53(1)(d) requires a public summary of the content used for training, on the Commission’s mandatory template. It lists data sources by type and, for scraped data, the top 10% of domains (for small and medium companies the top 5% or 1,000 domains, whichever is lower). It is updated every six months or on a material change.
Dates and fines
Models placed on the market after 2 August 2025 must comply at once, older models by 2 August 2027. The AI Office can enforce from 2 August 2026, with fines up to 3% of worldwide turnover or EUR 15 million, whichever is higher.
Code of Practice
The Commission’s voluntary Code of Practice has a copyright chapter. It asks signatories not to bypass paywalls, to use crawlers that follow robots.txt, to exclude sites recognised as persistently infringing, to limit infringing outputs and to name a contact point for complaints. Following the Code does not by itself prove compliance with copyright law.
These duties sit with the provider of the model. A company that only builds a third-party model into its product has other duties, see our EU AI Act checklist for IT companies and SaaS developers.
GDPR for personal data in training sets
Copyright clearance does not answer the personal data question. If a dataset contains information about identifiable people, the GDPR applies to collecting, storing and training.
Legal basis
Consent is rarely workable for data scraped at scale, so developers mostly rely on legitimate interest (Article 6(1)(f)). The European Data Protection Board (EDPB) applies a three-step test in its Opinion 28/2024: a real and lawful interest, necessity, and a balance against the rights of the people concerned, including what they reasonably expect.
Public does not mean expected
The EDPB draft Guidelines 03/2026 on web scraping for generative AI, adopted on 7 July 2026 for public consultation, say that a public post does not mean the author expects AI training. Technical limits such as robots.txt, login walls or CAPTCHAs make scraping less expected, and such sites should be excluded.
Sensitive data
Special categories under Article 9, such as health or political views, need an extra ground on top of the Article 6 basis. The Guidelines expect filters before collection and prompt deletion once such data is found.
Transparency
People must be informed. The exemption for disproportionate effort in Article 14(5)(b) should not be used routinely. The Guidelines expect a public notice naming sources and crawler details, a data protection impact assessment (DPIA) and measures against memorisation.
The EDPB assesses this case by case. It must be very unlikely that people can be identified or that their data can be extracted by queries. If a model was built on unlawfully processed data, that can affect whether using it is lawful, unless it is properly anonymised (Opinion 28/2024).
On 21 October 2026 the Council of the EU agreed its position on the Digital Omnibus. Among other things it would clarify the lawful ground for processing personal data when developing AI, with appropriate safeguards. This is a negotiating position only. Talks with the Parliament start once the Parliament adopts its own position, so the current GDPR text applies today.
How to prepare a dataset, step by step
A sequence that suits most AI projects.
List the sources
Record where each part of the data comes from: your own, licensed, users’ content, the public web or a purchase.
Start hereCheck the right to use
Read the licence or website terms and look for rights reservations in robots.txt and metadata. Keep dated evidence of what you checked.
Find the personal data
Decide whether people can be identified. If yes, choose and document the legal basis and run a DPIA.
Filter before training
Remove special-category data, sites with opt-outs, paywalled and pirate sources, and duplicates that make memorisation likelier.
Test the outputs
Check whether the model reproduces works or personal data, and add safeguards. Memorisation is where the Munich court found infringement.
Document and contract
If you are a provider, keep the copyright policy and the training summary. In contracts with data suppliers, ask for warranties on the origin of the data.
Collecting data for training and what the model later outputs are judged separately. The first depends on a licence or on the TDM exception and its opt-out, the second on whether the model memorises works. Personal data needs its own legal basis in both.
Collection and output are judged separately: the exception may cover the first and not the second.LegalBit team
Frequently asked questions
Can I scrape the web to train a model in the EU?
Possibly, if the Article 4 exception of the DSM Directive applies. You need lawful access, you must respect rights reservations made in a machine-readable way, and you may keep copies only as long as necessary. Personal data in the pages needs a GDPR basis as well.
What counts as a machine-readable opt-out?
The Directive speaks of machine-readable means, and the Commission’s Code of Practice points to robots.txt and other recognised protocols. The Hamburg appeal court held that a plain-language reservation on a website did not meet the requirement in the case before it. The point is disputed and a further appeal is open.
Does it matter where the training takes place?
For providers placing a general-purpose model on the EU market, recital 106 of the AI Act says the copyright duties apply regardless of where the training took place. The Munich court also applied German law to a model stored on servers in Germany.
Do I have to publish my training data?
No. Only providers of general-purpose AI models must publish a summary on the Commission’s template, and it describes the data sources, not the data itself.
Can I train on my users’ content?
Each layer needs its own ground: rights in the content, usually a licence in the terms, a GDPR legal basis and a clear notice to users. This is our assessment. If AI training is added to existing terms, do it openly and check the legal basis first.
Is training on lyrics or music already banned in Germany?
Not as such. The Munich court held against AI providers where the models memorised works and reproduced them. These are first-instance judgments and they are not final.
See also: AI-Generated Content and Copyright; EU AI Act checklist for IT companies and SaaS developers; A Marketer’s Guide to Fair Use.
- EUR-Lex: Directive (EU) 2019/790 on copyright in the Digital Single Market, Articles 3 and 4
- EUR-Lex: Regulation (EU) 2024/1689 (AI Act), Article 53 and recital 106
- EUR-Lex: Regulation (EU) 2016/679 (GDPR), Articles 6, 9 and 14
- European Commission: template for the summary of training content of general-purpose AI models
- European Commission: General-Purpose AI Code of Practice
- EDPB: Opinion 28/2024 on data protection aspects of AI models
- EDPB: Guidelines 03/2026 on web scraping in the context of generative AI (draft for public consultation)
- Hanseatic Higher Regional Court, Hamburg: press release on the judgment of 10 December 2025 (5 U 104/24)
- Munich Regional Court I: press release 11/2025 on the judgment in GEMA v OpenAI (42 O 14139/24)
- Munich Regional Court I: press release 16/2026 on the judgment in GEMA v Suno (42 O 763/25)
- EUR-Lex: Case C-250/25, Like Company, reference for a preliminary ruling
- Council of the EU: press release of 21 October 2026 on the position on the Digital Omnibus
This topic is our service
More on how we help AI and IT companies:
