AI and text and data mining clauses in library licences: a guide for academic librarians (2026)

Publishers are adding AI clauses to e-resource licences faster than most libraries can review them. This guide explains what EU law already guarantees for text and data mining by universities, where contracts can still restrict AI use, what library consortia recommend, and how to review and negotiate a licence clause by clause.

Published 9 October 2026 · Sources checked 9 October 2026

Also available in: Svenska · Nederlands · Norsk · Suomi · Dansk · Français · Español · Português

1. Why AI clauses are appearing now

Until 2023, most e-resource licences dealt with text and data mining in one or two sentences, usually allowing non-commercial research mining or pointing to a publisher API. Generative AI changed the commercial picture. Publishers now sell AI training licences, run their own AI research assistants, and want to stop subscribed content from reaching third-party models.

The International Coalition of Library Consortia (ICOLC) described the result in March 2024: e-resource vendors had begun presenting libraries and consortia with new licence terms restricting AI-related rights. In 2025 LIBER, the Association of European Research Libraries, said that it was seeing "license agreements which try to prevent AI and data mining activities in direct conflict with EU law".

The University of Turku Library told its users in April 2024 that "It is very likely that agreements for licensed resources will include more AI clauses as they are renewed." That expectation has been borne out. For a licensing librarian, the practical problem is that AI clauses arrive inside routine renewals, often with different wording from each publisher, and sometimes in website terms that the licence incorporates by reference.

2. What does Article 3 of the DSM Directive allow?

Article 3 of Directive (EU) 2019/790 lets research organisations and cultural heritage institutions make reproductions and extractions of works they have lawful access to, in order to carry out text and data mining for scientific research. A university, "including its libraries", is a research organisation (Art. 2(1)). Contract terms contrary to Article 3 are unenforceable (Art. 7(1)).

The key elements, as written in the Official Journal text:

• Definition. Text and data mining is "any automated analytical technique aimed at analysing text and data in digital form in order to generate information" (Art. 2(2)), including patterns, trends and correlations.
• Lawful access. Recital 14 says that where an institution has a subscription, "the persons attached thereto and covered by those subscriptions should be deemed to have lawful access." Freely available online content is also covered.
• Storage. Copies "shall be stored with an appropriate level of security and may be retained for the purposes of scientific research", including verification of results (Art. 3(2)).
• Publisher measures. Rightholders may apply measures to protect "the security and integrity of the networks and databases", but these "shall not go beyond what is necessary to achieve that objective" (Art. 3(3)).
• No contract override. "Any contractual provision contrary to the exceptions provided for in Articles 3, 5 and 6 shall be unenforceable." (Art. 7(1)).

Article 4(4) adds that the general exception "shall not affect the application of Article 3". A publisher's Article 4 opt-out therefore does not remove the research exception.

Validemic's analysis Article 3 protects the act of mining (reproduction and extraction for analysis). It does not, on its own terms, cover sharing licensed content with outside parties or making it available to others. That boundary is where most AI clause disputes sit.

3. How is Article 4 different?

Article 4 allows anyone with lawful access to mine works for any purpose, including commercial AI development. It applies only if the rightholder has not expressly reserved its use "in an appropriate manner, such as machine-readable means in the case of content made publicly available online" (Art. 4(3)). Article 7(1) does not list Article 4, so contracts can restrict it.

Recital 18 says that for online content it is only appropriate to reserve rights by machine-readable means, "including metadata and terms and conditions of a website or a service", and that in other cases reservation can be made by contractual agreements or a unilateral declaration.

Major publishers now signal reservations in several ways. What we found on the check date:

Validemic's analysis A machine-readable reservation expresses an Article 4 opt-out. Because Article 4(4) preserves Article 3, a reservation or consent requirement should not prevent mining for scientific research by a university with lawful access. Where a publisher's signals or terms suggest consent is needed even for research mining, ask the publisher to confirm in the licence that Article 3 uses are unaffected.

4. National transpositions in six countries

Member States transposed Articles 3 and 4 in their own words, and the details matter for licence review. We read the national texts below on the check date.

Two details are worth knowing when you read a licence. Sweden's 15 b § says copies may not be kept longer than necessary for the purpose, which is narrower wording than the Directive's retention rule. Finland's text expressly bars blocking research mining by technical means, which goes beyond the Directive's wording. Check the current consolidated text in your own country before relying on these summaries; the Danish consolidation we read lists a later amending act. We could not read the French transposition on Légifrance on the check date.

5. How does the AI Act refer to text and data mining?

The AI Act (Regulation (EU) 2024/1689) links to the DSM Directive in one place that matters for libraries. Article 53(1)(c) requires providers of general-purpose AI models to put in place a policy to comply with Union copyright law, and in particular to identify and comply with "a reservation of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790".

Three points follow for licensing librarians:

1. The duty falls on model providers, not on libraries. A university that uses an AI tool is normally a deployer, not a general-purpose model provider. Article 53(1)(d) also requires providers to publish a summary of training content, using a template from the AI Office. Article 53(2) keeps points (c) and (d) for open-source models.
2. Research mining remains distinct. Recital 105 notes that rightholders may reserve their rights to prevent text and data mining "unless this is done for the purposes of scientific research." Where rights are reserved, model providers "need to obtain an authorisation from rightsholders".
3. Reservations now have an enforcement path. The Commission's General-Purpose AI Code of Practice (copyright chapter) commits signatories to follow robots.txt and other appropriate machine-readable protocols, and not to circumvent restrictions "imposed by subscription models or paywalls". This is one reason publishers have formalised their Article 4 reservations.

The AI Omnibus (Regulation (EU) 2026/1744), in force since 27 July 2026, did not amend Article 53; we checked every amending point. For the wider timeline, see our AI Act explainer and AI Act guide for universities. The copyright text itself is on our DSM Copyright Directive page.

6. Text and data mining versus using content in AI tools

Researcher text and data mining and "AI use" are often merged in licence drafts, but they are legally different activities. Separating them is the most useful thing a reviewer can do, because the law protects some of them strongly and others not at all.

Validemic's analysis Uploading a licensed PDF to a consumer chatbot involves a copy held by a third party that may retain it and, depending on the plan, use it for training. Article 3(2) requires copies to be stored with an appropriate level of security. It is hard to argue that a consumer chatbot meets that test, so libraries should not present Article 3 as permission for this use. Enterprise or institutional AI services with contractual no-training and retention commitments are a different case, and that is where licence negotiation can make a real difference. Our guide on GDPR and AI tools in research covers the data protection side, and our fact sheets on ChatGPT, Microsoft Copilot and NotebookLM set out what each vendor documents about training.

7. Common restrictive clauses

Restrictive AI clauses fall into a small number of patterns. The examples below come from published sources, not from confidential agreements.

• Blanket bans on AI use. Elsevier's website terms say content may not be used in combination with an AI tool "(including to train an algorithm, test, process, analyse, generate output and/or develop any form of" AI), unless expressly permitted, for example by a licence or subscription agreement.
• No training, including indirect training. The Authors Alliance analysed a clause proposed by a commercial publisher to a university that barred users from using content to "directly or indirectly develop, train, program, improve, and/or enrich any artificial intelligence tool".
• No uploading to third-party AI tools. The same clause allowed content to be passed to a third-party AI tool only if the tool "is not trained or fine-tuned using the Content". The University of Turku reports that FinELib's Springer Compact terms (2024 to 2025) say material "must also not be used directly or indirectly to train artificial intelligence."
• Hosting conditions. Turku summarises the FinELib Elsevier agreement (2024 to 2025) as requiring AI tools to run in the subscriber's own environment, or, if hosted by a third party, to keep uploaded material available only to the subscriber and authorised users.
• Restrictions on software, not just use. Teresa Nobre of COMMUNIA, writing on the Kluwer Copyright Blog in May 2025, quoted licence terms reported by libraries that targeted "the use of browsers with integrated or connected functionality of an automated IT application".
• Fees for rights the law already grants. The University of California's Office of Scholarly Communication says some vendors "attempt to preclude, or charge additional and unsupportable fees for, fair uses like training AI tools".
• New liability and monitoring duties. Clauses that make the institution liable for every user's AI use, or require it to monitor users. ICOLC's statement asks that AI clauses "not introduce new liability clauses beyond the scope of what has previously been agreed upon".

Nobre's post also raises a structural risk: many licences choose the law of a non-EU country. Article 7(1) protects Article 3 under EU law, but a foreign governing law clause can make enforcement harder. The post suggests treating the research exception as an overriding mandatory provision. In practice, the simpler step is to ask for the law of an EU Member State, or an express statement that statutory exceptions are not affected.

Reviewing a vendor right now? Validemic checks the vendor's documents against GDPR and the EU AI Act and cites every finding. Try the demo workspace

8. Model language from library bodies

Several library organisations have published positions you can cite in negotiations. We read the following on the check date.

ICOLC Statement on AI in Licensing (22 March 2024)

The statement sets out principles for AI clauses in consortium licences. In summary, AI clauses should:

• "permit the use of AI for any and all legal purposes" supporting non-commercial research, teaching and learning;
• in multi-year licences, provide formal opportunities for revision as law and technology change;
• not restrict user actions "in ways that are fundamentally unenforceable";
• not prevent authorised users from making licensed content fully accessible;
• not be introduced in a way that limits a consortium's normal review;
• not add new liability beyond what was previously agreed.

LIBER endorsed the statement in 2025, asking vendors to follow its principles so that researchers can exercise their rights under EU law. IReL, the Irish research e-library consortium, endorsed it in May 2024 and said it is "committed to maintaining the rights of our users and the values of members in our agreements with vendors."

University of California (2024)

UC Libraries published a position in March 2024 and adaptable licensing language. The UC Office of Scholarly Communication summarises that licensed products may be used with AI tools as long as doing so would not:

• "create a competing or commercial product or service for use by third parties";
• "unreasonably disrupt the functionality of the subscribed products";
• "reproduce or redistribute the subscribed products for third parties".

The post also recommends that licences "require commercially reasonable security measures", and notes that in "more than forty countries" publishers cannot use contracts to override copyright exceptions for research. It says: "we hope to see express contractual allowance for AI training become the norm in academic licensing".

LIBER

LIBER campaigned for a text and data mining exception before the DSM Directive and in March 2019 welcomed "a Europe-wide copyright exception for Text and Data Mining". Its 2025 endorsement of the ICOLC statement is the most direct position on AI licence clauses from LIBER that we found.

Other bodies

Jisc in the UK has published guidance on resisting restrictive AI clauses in licences through its National Centre for AI, but the page returned an access error when we tried to read it on the check date, so we do not summarise it here. In the United States, SPARC published an analysis in March 2024 of publisher website footers that reserve AI rights, recommending fair use savings clauses and making the negotiated licence take precedence over click-through terms. We did not find published AI model clauses from Bibsam (Sweden), UKB or SURF (the Netherlands), Couperin (France) or SPARC Europe on the check date. Consortia often share positions with members only, so ask your consortium directly.

9. What university libraries tell their users

Library user guidance is a good signal of what licences actually allow, because libraries write it after reading their agreements. The pages below, from six countries, show both the restrictive default and the direction of travel.

"Det är inte tillåtet att [...] använda upphovsrättsskyddade publikationer och annat material som inmatning i AI-verktyg." (It is not permitted to use copyrighted publications and other material as input to AI tools.)

"Het is niet toegestaan gelicentieerd materiaal te gebruiken in AI tools, bijvoorbeeld voor maken van een samenvatting." (It is not permitted to use licensed material in AI tools, for example to make a summary.)

"Please note, you are not allowed to upload copyrighted materials to an AI tool."

SDU adds that users are "free to upload patents and open access publications". Helsinki warns that a breach of the conditions of use can lead to all users losing the right to use the resource.

"bulk uploading subscription-based content to third-party LLMs is not advisable."

Queen's University Belfast also points out that even openly licensed content can be a problem: under a Creative Commons licence that restricts commercial use, "sharing this content with an AI could violate the terms, as many AI platforms operate commercially." The University of East Anglia reminds staff that the UK exception for non-commercial research text and data mining "also trumps any publisher's contract restrictions" where the library already subscribes.

"The University endeavors to secure AI usage and training rights across all of our e-resource licenses."

Berkeley adds that "licenses may not allow for use or training of their content with public versions of AI tools" and asks users to contact the library before using AI with any licensed resource. The University of Waterloo (Canada) says "Most library resources can't be shared with or uploaded to generative artificial intelligence (GenAI) tools" and that breaches "can result in loss of access for individual users or the entire campus." American University (United States) applies its restriction to "all AI tools, including AU-licensed platforms", but not to open access material or to AI search features embedded in library databases such as JSTOR.

The University of Turku Library (Finland) takes a more granular approach, explaining agreement by agreement what the FinELib Elsevier and Springer Compact terms allow, and pointing users to resource-specific terms. Turku reports that the Elsevier agreement permits AI-assisted abbreviations, edits and translations for an authorised user's personal use, under hosting conditions.

Validemic's analysis Blanket "no AI" user guidance is simple to communicate, but it goes further than many licences, and it does not mention the Article 3 research exception. A tiered message works better: research text and data mining is protected by law; uploading to public AI tools is not allowed unless the licence says so; approved institutional AI tools may be allowed for specific resources. For examples of which AI tools European universities approve, see AI tools approved at European universities and Nordic university AI services.

10. Clause, risk and suggested position

Use this table when reading a draft. The suggested positions reflect the ICOLC principles and the UC language above, combined with the EU legal position; adapt them to your consortium's mandate.

Our licence AI clause checker flags these patterns automatically and suggests which position from this table applies.

11. Negotiation checklist

12. Privacy in publisher AI tools

Publisher AI assistants are licensed like content, but they work like software services. They process user accounts, prompts, conversation history and usage logs, and they send prompts to large language model providers. That makes them a GDPR question as well as a licensing one, and it is easy to miss because the agreement is negotiated by the library rather than by IT procurement.

What vendors say publicly (checked 9 October 2026):

• Scopus AI (Elsevier). Elsevier says OpenAI's ChatGPT, hosted on Microsoft Azure, is among the models it uses, with an agreement that "no user queries will be stored or used to train or improve ChatGPT". Its product page describes standard conversation history as kept until the contract ends, and temporary conversations for up to 30 days. The support centre advises: "Do not enter personal, confidential, or sensitive information into Scopus AI." Our Scopus AI fact sheet sets out the details.
• Clarivate (Web of Science, ProQuest and Primo research assistants). Clarivate's AI policy says: "Clarivate does not use your data to directly or indirectly train LLMs." We did not find the model provider named on that page.
• JSTOR research tool (ITHAKA). ITHAKA's 2024 promotional material says: "Your prompts and content text are never used to further train models and are deleted after 30 days." JSTOR's support documentation says administrators can enable or disable the tool for all users at an institution.

Questions to ask before switching a publisher AI feature on:

1. Is the publisher a processor for the university, or a controller of user data? Is there a data processing agreement covering the AI feature?
2. Which AI model providers receive prompts, and in which regions? Are they listed as subprocessors?
3. How long are prompts, answers and conversation history kept, and can the institution shorten this?
4. Are prompts or usage data used to train or improve the publisher's own models, not only the third-party model?
5. Can the feature be switched off per institution, and is it off or on by default?
6. Are prompts linked to identifiable users in usage reports or logs shared with the institution?

Researchers may type unpublished hypotheses, draft text or participant details into these tools. The library's user guidance should say what not to enter. Our guide on library patron privacy and GDPR covers user data in library systems more broadly, and the vendor assessment guide explains the full review.

13. A step-by-step review process

1. Collect the full text. Get the licence, all schedules, and every document it incorporates by reference, including website terms and AI policies. Note the date of each.
2. Run a first scan. Paste the text into the licence AI clause checker or search for terms such as "artificial intelligence", "machine learning", "train", "mining", "automated", "robot", "third party" and "derivative".
3. Separate the activities. For each flagged clause, decide whether it touches research mining (Art. 3), general mining (Art. 4), uploading to AI tools, model training, or the publisher's own AI feature. Use the table in section 6.
4. Check the law that applies. Confirm the governing law and your national transposition. Note whether the clause would be unenforceable under it.
5. Compare with your position. Use the clause table and your consortium's mandate. Many consortia have adopted or endorsed the ICOLC principles.
6. Review the AI feature as a vendor. If the platform includes an AI assistant, ask the privacy questions above and involve the DPO. Our DPA checker helps with the data processing agreement.
7. Negotiate and document. Send marked-up wording, not only comments. Record what was accepted, what was rejected, and why.
8. Tell users. Publish per-resource AI permissions in the catalogue or ERM, and update general guidance so that it reflects the research exception.
9. Monitor. Website terms, AI policies and subprocessor lists change during the licence term. Track them and review at renewal. Our subprocessor alerts tool can help with the AI provider side.

If your institution also has a general AI policy, link licence permissions to it, so that users see one consistent rule. Our AI policy template for universities can be a starting point.

Sources

1. Directive (EU) 2019/790 on copyright and related rights in the Digital Single Market, Articles 2, 3, 4 and 7, recitals 14 and 18, read in the Official Journal text via publications.europa.eu (retrieved 9 October 2026).
2. Regulation (EU) 2024/1689 (AI Act), Article 53 and recitals 105 and 106 (retrieved 9 October 2026).
3. Regulation (EU) 2026/1744 (Digital Omnibus on AI), Article 1 (retrieved 9 October 2026).
4. European Commission: The General-Purpose AI Code of Practice, copyright chapter (retrieved 9 October 2026).
5. Sweden: Lag (1960:729) om upphovsrätt, 15 a to 15 c §§ (retrieved 9 October 2026).
6. Netherlands: Auteurswet, articles 15n and 15o (retrieved 9 October 2026).
7. Germany: UrhG § 44b, § 60d and § 60g (retrieved 9 October 2026).
8. Finland: Tekijänoikeuslaki (404/1961), 13 b § (retrieved 9 October 2026).
9. Denmark: Ophavsretsloven (LBK nr 1093 af 20/08/2023), §§ 11 b and 11 c (retrieved 9 October 2026).
10. United Kingdom: Copyright, Designs and Patents Act 1988, section 29A (retrieved 9 October 2026).
11. ICOLC Statement on AI in Licensing, 22 March 2024 (retrieved 9 October 2026).
12. LIBER endorses ICOLC Statement on AI in Licensing, 2025 (retrieved 9 October 2026).
13. IReL endorses the ICOLC Statement on AI in Licensing, 16 May 2024 (retrieved 9 October 2026).
14. LIBER welcomes copyright improvements for libraries, 26 March 2019 (retrieved 9 October 2026).
15. University of California Office of Scholarly Communication: Fair use rights to conduct text and data mining and use artificial intelligence tools are essential for UC research and teaching, 12 March 2024 (retrieved 9 October 2026).
16. UC Berkeley Library: Conditions of use for electronic resources (retrieved 9 October 2026).
17. University of Helsinki Library: Villkor för användning av e-resurser (retrieved 9 October 2026).
18. University of Turku: Electronic resources and use of artificial intelligence tools, 23 April 2024 (retrieved 9 October 2026).
19. Radboud University: Gebruiksvoorwaarden digitale bibliotheek (retrieved 9 October 2026).
20. University of Southern Denmark Library: Copyright (science and technology guide) (retrieved 9 October 2026).
21. Queen's University Belfast Library: Uploading to LLMs (retrieved 9 October 2026).
22. University of East Anglia Library: Using library online resources (retrieved 9 October 2026).
23. University of Waterloo Library: Use of library resources with AI (retrieved 9 October 2026).
24. American University Library: May I upload library electronic resources into an AI tool or LLM? (retrieved 9 October 2026).
25. Elsevier website terms and conditions, revised 12 March 2026, and Elsevier TDMRep licence (retrieved 9 October 2026).
26. Wiley Online Library TDM reservation file and wiley.com footer (retrieved 9 October 2026).
27. Springer Nature: TDM reservation policy (retrieved 9 October 2026).
28. Sage: TDM and AI policy (retrieved 9 October 2026).
29. Taylor & Francis: Text and data mining policy (retrieved 9 October 2026).
30. Elsevier: Scopus AI and Scopus AI support article (retrieved 9 October 2026).
31. Clarivate: AI Policy (academia) (retrieved 9 October 2026).
32. ITHAKA: JSTOR research tool handout (2024) and JSTOR support: managing institutional access to the AI research tool (retrieved 9 October 2026).
33. Authors Alliance (Samberg and Hansen): Restricting innovation: how publisher contracts undermine scholarly AI research, 6 December 2024 (retrieved 9 October 2026).
34. Kluwer Copyright Blog (Teresa Nobre): Unfair licensing practices: the library experience, 15 May 2025 (retrieved 9 October 2026).
35. SPARC (Kyle Courtney): All TDM and AI rights reserved?, 28 March 2024 (retrieved 9 October 2026).

About this page

Sources checked on 9 October 2026. We read the DSM Directive, the AI Act and the AI Omnibus in the Official Journal versions published by the EU Publications Office, the national statutes on official government sites, and the library, consortium and publisher pages listed above. Quotes in other languages are followed by our own English translation. We could not read the French transposition on Légifrance or Jisc's guidance on AI clauses on the check date, so we do not describe them. Where we give our own interpretation, it is labelled as Validemic's analysis. This page is general information, not legal advice, and it does not describe the terms of any confidential licence. If you spot an error, or your consortium has published AI licence principles we should include, please contact us and we will correct it.