DataPacts

EU Database Right (Directive 96/9/EC) and Licensing Datasets for AI Training

Checked

At a glance

The EU database right is the sui generis right of Directive 96/9/EC that lets the maker of a database built with substantial investment prevent extraction or re-utilization of the whole or a substantial part of its contents for 15 years, and it is the right a company licenses when it licenses a dataset rather than individual works.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

The Directive was adopted on March 11, 1996 (OJ L 77, 27.3.1996, p. 20; CELEX 31996L0009; EUR-Lex text).

Member states had to transpose it before January 1, 1998 (Art. 16(1)), and it is in force as of 2026-11-17. In AI training data licensing it is the database layer, deciding whether a dataset is an asset a maker can license.

The Court of Justice reads the qualifying investment narrowly for created data. Copying a substantial part into a training corpus is extraction, the term is 15 years, and the right is licensable under Art. 7(3).

The DSM Directive's mining exceptions and Data Act Art. 43 cut it back, and no ruling on AI training is recorded.

The Directive's Sui Generis Provisions at a Glance

Ten provisions of Directive 96/9/EC and two of later EU acts decide whether a dataset carries the right and what a license or an AI developer can do with it. The table is as of 2026-11-17, and the sections below read each provision.

Provision Rule What it means for a training dataset Source
Art. 1(2) A database is a collection of independent works, data or other materials, arranged systematically or methodically and individually accessible The test for whether a dataset is a database at all eur-lex.europa.eu
Art. 3(1)-(2) Copyright for an original selection or arrangement; it does not extend to the contents The structure is protected, the data is not eur-lex.europa.eu
Art. 7(1) A maker with substantial investment in obtaining, verification or presentation may prevent extraction or re-utilization of the whole or a substantial part The core right eur-lex.europa.eu
Art. 7(2) Extraction is transfer to another medium; re-utilization is making available to the public Copying a substantial part into a corpus is extraction eur-lex.europa.eu
Art. 7(3) The right may be transferred, assigned or licensed The licensable asset eur-lex.europa.eu
Art. 7(5) Repeated and systematic extraction of insubstantial parts is barred where it harms normal exploitation No piecemeal scraping around the rule eur-lex.europa.eu
Art. 8(1) A lawful user may extract insubstantial parts for any purpose Small samples stay free eur-lex.europa.eu
Art. 10(1)-(3) 15 years from 1 January after completion or first making available; a new term on substantial new investment Updated archives can renew eur-lex.europa.eu
Art. 11(1)-(2) Makers or rightholders who are EU nationals, EU residents or EU-established companies Who can claim it eur-lex.europa.eu
Art. 15 Terms contrary to Arts. 6(1) and 8 are null and void Licenses cannot remove lawful-user rights eur-lex.europa.eu
DSM Arts. 3(1), 4(1), 4(3) Mining exceptions to Art. 7(1); Art. 4 applies unless reserved Mining a lawful copy may need no license eur-lex.europa.eu (CELEX 32019L0790)
Data Act Art. 43 No sui generis right for data from in-scope connected products or related services Product data sits outside the right eur-lex.europa.eu (CELEX 32023R2854)

When a Dataset Qualifies: Substantial Investment in Obtaining, Verifying or Presenting Contents

A dataset carries the sui generis right only if it is a database in the Directive's sense and its maker shows a qualitatively or quantitatively substantial investment in obtaining, verifying or presenting its contents (Art. 7(1)).

Art. 1(2) defines a database as a collection of independent works, data or other materials, arranged systematically or methodically and individually accessible. A collection of support tickets, product records or labeled images can meet that definition, while a single document cannot.

The right applies irrespective of whether the database or its contents qualify for copyright (Art. 7(4)).

Recital 40 counts money, time, effort or energy as investment, so the build records of who paid for collection, checking and presentation are the evidence a licensor needs, and they belong in the rights audit before licensing data to AI companies.

Created Data Does Not Count: British Horseracing Board (C-203/02)

The Court of Justice held on November 9, 2004 that investment in obtaining contents means resources used to seek out existing independent materials and collect them, not resources used to create the materials (C-203/02, British Horseracing Board v William Hill, Grand Chamber).

Verification means monitoring the accuracy of materials collected, not checks made while creating them, and the resources used to draw up a list of horses in a race did not qualify.

For AI data, records a company creates through its own operations may struggle to meet this test, while a curated collection of existing materials fits it better. Public accessibility does not end the right.

Who Holds the Right: The Maker Who Took the Investment Risk

The right belongs to the maker, the person who takes the initiative and the risk of investing, so a subcontractor is not the maker (recital 41).

The right covers makers or rightholders who are EU nationals or habitually resident in the EU, and companies formed under member-state law with their registered office, central administration or principal place of business in the EU (a registered office alone needs a genuine ongoing link) (Art. 11(1)-(2)). Third-country databases are covered only through Council agreements (Art. 11(3)).

Extraction and Re-utilization: What Copying a Dataset for Training Touches

Extraction is the permanent or temporary transfer of all or a substantial part of a database's contents to another medium, and re-utilization is making them available to the public, so copying a database into a training corpus is the kind of act the right controls (Art. 7(2)).

Getty Images v Stability AI, an AI training case that included a database claim, never tested it: the claim was abandoned before the judgment of November 4, 2025 ([2025] EWHC 2863 (Ch)), under UK law (see Getty Images v. Stability AI in the UK High Court).

Only a substantial part counts, judged by volume against the whole or by the investment behind the part taken (C-203/02).

Repeated and systematic extraction of insubstantial parts is barred where it conflicts with normal exploitation or unreasonably prejudices the maker (Art. 7(5)), which the Court reads as acts that cumulatively rebuild a substantial part. A lawful user of a database made available to the public may extract insubstantial parts for any purpose (Art. 8(1)).

Copying a Freely Accessible Database: CV-Online (C-762/19)

On June 3, 2021 the Court of Justice ruled that a specialized search engine which copies and indexes the whole or a substantial part of a database freely accessible on the internet extracts and re-utilizes it (C-762/19, CV-Online Latvia v Melons, Fifth Chamber).

The maker may prohibit this only where the acts adversely affect its investment, putting at risk the possibility of redeeming it through the database's normal operation, which the national court verifies.

The Court sought a fair balance between makers and the users and competitors who build innovative products from the information (para. 41). The ruling concerns a job-ad search engine, not training, and no decision applying it to training is recorded. The other claims are covered under whether web scraping for AI training is legal.

How Long the Right Lasts: 15 Years, Renewed by Substantial New Investment

The right runs from completion of the database and expires 15 years from January 1 of the year after completion, or 15 years from January 1 after first making available to the public if that happens within the first term (Art. 10(1)-(2)).

  • Art. 10(1): the term starts at completion.
  • Art. 10(2): first making available to the public within that term starts a new count from the following January 1.
  • Art. 10(3): a substantial change, including accumulated additions, deletions or alterations, that amounts to substantial new investment gives the resulting database its own term.

A continuously updated archive can therefore stay protected, and the maker must prove the completion date and the new investment (recitals 53 and 54).

Licensing the Right Under Article 7(3)

The sui generis right may be transferred, assigned or granted under contractual license (Art. 7(3)), so a dataset license for AI training can permit the acts the right controls.

A grant covers extraction and re-utilization of the whole or substantial parts, not copyright in the works inside, which stays with their rightholders (Art. 7(4)). ODC-BY 1.0 governs only rights over the database, not its contents (s. 2.4), as shown in dataset licenses for AI training.

A license for this right states:

  • which acts it permits: extraction for training, re-utilization by publishing outputs, or sublicensing;
  • which parts of the dataset it reaches;
  • how long it runs relative to the 15-year term.

Art. 15 voids contract terms contrary to the lawful-user rights of Arts. 6(1) and 8, so a license cannot forbid a lawful user from taking insubstantial parts.

In Ryanair v PR Aviation (C-30/14, decided January 15, 2015) the Court held that where a database has neither copyright nor the sui generis right, the Directive does not apply and its author may set contractual limits, without prejudice to national law.

Protected or not, the contract does most of the work, as the clauses of an AI data licensing agreement show.

Where the Right Stops for AI Training Data

Two later EU acts cut the database right back where AI training data is concerned: the DSM Directive's text and data mining exceptions and Article 43 of the Data Act.

Mining Exceptions That Reach Article 7(1)

Directive (EU) 2019/790 requires member states to except reproductions and extractions for text and data mining from the Art. 7(1) right, for research organizations and cultural heritage institutions (Art. 3(1)) and for anyone with lawful access (Art. 4(1)).

Art. 4 applies only if the rightholder has not expressly reserved the use, machine-readably for content made available online (Art. 4(3)). DSM Art. 24(1) made the Directive's research exceptions (Arts. 6(2)(b) and 9(b)) apply without prejudice to them.

Spain's RDL 24/2021 art. 67(6)-(7), for example, extends its exception to substantial parts of databases. Reservation and contract override are covered under the EU text and data mining exception and the opt-out.

No Database Right in Connected-Product Data: Article 43 of Regulation (EU) 2023/2854

The sui generis right of Art. 7 of Directive 96/9/EC does not apply when data is obtained from or generated by a connected product or related service within the Data Act's scope, in particular in relation to its Articles 4 and 5 (Art. 43). The Data Act applies from 2025-09-12 (Art. 50).

Recital 112 states the aim: holders of sensor and other machine-generated data should not use the right to block users' access and sharing rights, though it can still apply to databases with data outside the Regulation's scope if Art. 7(1) is met.

A manufacturer therefore cannot rely on the database right to license or withhold product data, and what it may license is governed by the user contract under the EU Data Act rules on product data.

A single dataset can carry up to three separate protections at once: copyright in its structure, the sui generis right in its contents as a whole, and the rights in each work it contains.

  • Copyright protects an original selection or arrangement, not the contents (Art. 3(1)-(2)).
  • The sui generis right applies irrespective of copyright in the database or its contents (Art. 7(4)).
  • The Directive is without prejudice to trade secrets, confidentiality, data protection and the law of contract (Art. 13).

Confidentiality of a delivered dataset falls under Directive (EU) 2016/943, covered in EU Trade Secrets Directive rules for licensed datasets.

Relying on the Database Right When Licensing a Dataset: Benefits, Risks, Downsides and Alternatives, Including Not Licensing

The database right gives a dataset maker an asset it can license under Art. 7(3), but its reach is narrow, and licensing on contract alone or not licensing at all remain valid choices.

Benefits

  • It applies to every qualifying database, irrespective of copyright (Arts. 7(1), 7(4)).
  • It is licensable and transferable (Art. 7(3)).
  • Substantial new investment renews it (Art. 10(3)).

Risks

  • Data the company created itself may not qualify (C-203/02).
  • Others may mine a lawfully accessible copy unless the use is reserved (DSM Art. 4).
  • Connected-product data is excluded (Data Act Art. 43).
  • No ruling on AI training is recorded.

Downsides

  • The maker must prove its investment and the dates (recitals 41, 53, 54).
  • Insubstantial parts stay free for lawful users (Arts. 8, 15).
  • Protection depends on national transposition and on EU nationality or establishment (Art. 11).

Alternatives

  • License on contract terms and confidentiality instead (C-30/14, Directive (EU) 2016/943).
  • License individual works through their rightholders.
  • Publish under an open dataset license.
  • Not licensing at all.

Every option is covered in the alternatives to selling data to AI labs, including not selling.

Beyond This Page: Other EU Instruments, Other Rights and Seller Steps

The database right is one of the EU instruments a training data deal meets. The parent page sets them side by side under EU rules for AI training data.

How This Page Is Sourced and Dated

Every EU rule here cites the Official Journal text of Directive 96/9/EC, a later EU act or a Court of Justice judgment, with the date it was read.

The method is described under how we source, date and check every figure, and this page is general information about the EU database right, not legal advice on any specific dataset or license.

Frequently Asked Questions

What is the sui generis database right?

The sui generis database right is the EU right of a database maker who made a substantial investment in obtaining, verifying or presenting contents to prevent extraction or re-utilization of a substantial part (Directive 96/9/EC, Art. 7(1)).

Copyright protects a database's original selection or arrangement, not the data in it (Art. 3), while the sui generis right protects the investment in the contents, separately (Art. 7(4)).

How long does the EU database right last?

It lasts 15 years from January 1 after completion or first making available, with a new term after substantial new investment (Art. 10).

Does the database right stop AI training on a dataset?

Not as such: copying a substantial part needs a license or an exception, and the DSM Directive's mining exceptions cover Art. 7(1) acts, Art. 4 unless reserved. No ruling applying the right to AI training is recorded.