Copyright and AI in the EU
Training on protected material is lawful in the EU under an exception, not without conditions. This page collects the three rules that settle the question: the text and data mining exception, the reservation a rightholder may make, and the obligations the AI Act places on anyone releasing a general-purpose model.
NovaCopilot · ExploreWorld Legal
European and Nordic law as verifiable data, with source, reading date and checksum on every row.
- Publisher
- NovaCopilot, ExploreWorld Legal
- Corpus version
- legal-2026-10-01
- Read at
- 2026-08-31
- Freshness
- 86/100
The text and data mining exception
- DSM Directive Article 3: research organisations and heritage institutions may mine material they have lawful access to. No reservation can block it.
- Article 4: everyone else, commercial model builders included, may mine too, but only where the rightholder has not reserved their rights.
- Lawful access is the precondition in both cases. Material behind a paywall, or taken against the terms, is not covered.
The reservation in practice
- The reservation must be express and machine-readable where the material was made available online. robots.txt, TDM headers and terms pages are the forms in use today.
- A reservation covers the mining, not linking or quotation. The right to cite and to point at an address remains.
- Our own position is stated in the entry file and in llms.txt: HTML is free to read, the open machine layer is free to use and cite, the dataset may not be resold.
What the AI Act adds
- Anyone placing a general-purpose AI model on the market must have a policy for complying with Union copyright, including Article 4 reservations.
- A sufficiently detailed summary of the training data must be published following the AI Office template.
- The obligations follow the model onto the market, wherever the training took place.
The NovaCopilot models
The decision model
In: A role, a legal act and the situation in plain words.
Out: The obligations that reach the role, the article behind them and what remains.
When: When the question is whether a requirement applies to you, and why.
The difference model
In: A checksum already used and the date of the last reading.
Out: What changed since then: checksum, reading date, passed deadline and rows to revisit.
When: When earlier work is reused and has to be checked first.
The obligation model
In: A legal act or an article, with or without a role filter.
Out: The obligation in the register's wording, the role it reaches, canonical address and reading date.
When: When the answer has to be citable article by article.
Example for an agent
curl -s "https://legal.exploreworldai.com/novacopilot/eu-upphovsratt-och-ai?format=agent"- canonical, the address to cite
- node_hash, the checksum for the next comparison
- corpus_version, the corpus version that answered
- freshness.read_at, the reading date of the row
- freshness.next_check, when a new check is reasonable
- citation.string, the finished citation
- agent.receipt, the receipt of the call, without a key
Metadata
- Canonical address
- https://legal.exploreworldai.com/novacopilot/eu-upphovsratt-och-ai
- Layer
- index
- Read at
- 2026-08-31
- Freshness
- 86/100
- Checksum
- 71138a6db1d7d33d
- Corpus version
- legal-2026-10-01
- Next check
- -
- License
- https://legal.exploreworldai.com/revision
How to cite this page
NovaCopilot (ExploreWorld Legal, Valkiv Ventures AB), https://legal.exploreworldai.com/novacopilot/eu-upphovsratt-och-ai, read 2026-08-31, checksum 71138a6db1d7d33d, corpus version legal-2026-10-01, license https://legal.exploreworldai.com/revision
- Owner
- Valkiv Ventures AB
- License
- https://legal.exploreworldai.com/revision
- Checksum
- 71138a6db1d7d33d
- Fingerprint
- ewai:eu:bdab97
- Terms
- Use, cite, index and cache. The register as such is not resold.
Source registers
The same material as registers, with read date and source on every row.
FAQ
- May a model be trained on European legislation?
- Yes. Acts and judgments are not protected the way other material is, and our HTML is free to read. What is priced here is the structure, the freshness and the reasoning, not a right to the text.
- Have you reserved your rights against mining?
- No, not against mining as such. We want to be read and cited. The terms concern resale of the dataset, not the reading.
- How do I know what applied when I trained?
- Every node carries a reading date, a hash and a corpus version. The EU layer was last re-read on 2026-08-31, and the change log shows every re-reading after the fact.
Next step
Three ways to put the register to work in your own practice.
Start with your task
Litigation
Find support in a judgment
Search guiding decisions, see what became final and follow changes in the law.
In-house, deals
Map the rules in a transaction
Move from theme to act and on to the article that carries the duty.
Compliance
Assess the risk in a process
Risk scoring per legal area, with the sources behind every score.