Database Rights and AI Training on EU Databases

A European dataset can be free of copyright and still not free to train on. The sui generis database right is a separate property right with its own subsistence test, its own infringement standard and its own term, and a training pipeline cleared for copyright has not been cleared for it.

A second right over the same data

Directive 96/9/EC on the legal protection of databases creates two regimes. Chapter II gives copyright protection to a database whose selection or arrangement of contents constitutes the author’s own intellectual creation. Chapter III creates something else entirely: a sui generis right, in Article 7, for the maker of a database who shows substantial investment in obtaining, verifying or presenting the contents. The Directive is at EUR-Lex.

The sui generis right has no equivalent in US law, which is why it is routinely missed by teams whose legal instincts were formed there. It protects investment rather than creativity, so it can attach to a plainly unoriginal compilation — an exhaustive list, a set of measurements, a price feed — that copyright would not touch at all. Its term is fifteen years from completion under Article 10, and substantial new investment in the database restarts that term for the resulting database, which in practice means a continuously maintained database can hold the right indefinitely.

When the right subsists

The most useful thing to know is the limit the Court of Justice put on it. In a group of judgments delivered on 9 November 2004 — The British Horseracing Board v William Hill, C-203/02, and the Fixtures Marketing cases C-46/02, C-338/02 and C-444/02 — the Court held that investment in obtaining the contents means investment in seeking out and collecting existing independent materials, and does not include the resources used to create those materials in the first place.

That distinction has real consequences for AI training. A body that generates the data as a by-product of its own activity — a league that creates the fixture list, a racing authority that creates the race data — cannot count the cost of creating it toward the substantial investment test. A body that spends money collecting data produced elsewhere can. So a public register compiled from third-party filings is a stronger candidate for protection than an operator’s own transaction log, which is the opposite of most people’s intuition. The judgments are searchable on the Court of Justice’s case-law database.

What counts as extraction

Article 7(1) gives the maker the right to prevent extraction or re-utilisation of the whole or of a substantial part of the contents, evaluated qualitatively or quantitatively. Article 7(5) adds the anti-circumvention rule that matters most to crawlers: repeated and systematic extraction of insubstantial parts is prohibited where it conflicts with normal exploitation or unreasonably prejudices the maker’s legitimate interests.

A polite crawler that takes a few rows per request over months is exactly the pattern Article 7(5) is aimed at. Rate limiting your scraper is good manners and good engineering; it is not a defence, because the provision exists precisely to stop a substantial taking being assembled out of insubstantial ones. “Extraction” is also defined broadly as permanent or temporary transfer to another medium by any means, so copying into a training corpus is squarely within it even if the corpus is never published.

Whether the TDM exceptions cover it

Here the answer is better than the fear. Directive (EU) 2019/790, the CDSM Directive, created two text-and-data-mining exceptions, and both were drafted to reach the database right as well as copyright. Article 3, the exception for research organisations and cultural heritage institutions, and Article 4, the general exception subject to an opt-out, each apply expressly to the sui generis right in Article 7(1) of Directive 96/9/EC alongside the copyright reproductions they cover. The text is at EUR-Lex.

Two conditions come with that, and both are load-bearing:

• Lawful access. Both exceptions require it. Content behind a paywall you did not pay for, or behind terms of use you breached to reach, is not lawfully accessed, and the exception does not begin to apply. This is the single most common way a European TDM defence fails.
• The Article 4(3) reservation. The general exception does not apply where the rightholder has expressly reserved the use in an appropriate manner — machine-readable, for content made publicly available online. A database maker may reserve, and the reservation defeats Article 4 for the database right just as it does for copyright.

What is not settled is how far Article 3’s research carve-out stretches when a research organisation’s output is later commercialised, and whether a reservation expressed only in natural-language terms of service is “appropriate”. Neither question has been answered by the Court of Justice, and views among national courts and commentators differ. See the Article 4 opt-out for the mechanics of the reservation itself.

What the Data Act changed

One narrowing arrived recently and is easy to miss. Regulation (EU) 2023/2854, the Data Act, provides in Article 43 that the sui generis right does not apply to databases containing data obtained from or generated by the use of a connected product or related service. The purpose is to stop the database right being used to block the access and portability rights the Data Act creates over device-generated data, and the relevant provisions of the Regulation became applicable from 12 September 2025. The text is at EUR-Lex.

For an AI team the practical reading is that IoT and telemetry data has become materially easier to work with, and that the database right remains fully in play for everything else — public registers, curated corpora, bibliographic and financial data, structured web sources. See connected-device data under the Data Act for how the access rights themselves work, and training data copyright law for the copyright layer sitting over the same material.