DataPacts

Libraries, Archives and Museums Licensing Collections for AI

Checked

At a glance

Libraries, archives and museums license collections to AI developers mostly by releasing digitized holdings under their own access terms, and the one priced training license with a library on public record runs the other way: the library pays.

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 Institutional Data Initiative at Harvard Law School Library requires contact for commercial use of Institutional Books 1.0, including AI model training (checked November 9, 2026), and the National Library of Norway pays Kopinor NOK 45,000,000 for 2026 for press training rights.

Copyright status, third-party rights, public-sector rules and personal data decide what a collection can license.

Collections are one source in AI training data licensing, no library, archive or museum has a published direct license with a price, and leaving a collection unlicensed is a valid outcome.

DataPacts Editorial Team checked every figure here against its primary source; the page on how we source, date and check every figure sets out the rules.

What Libraries, Archives and Museums Hold That AI Developers Want

Memory institutions hold four kinds of material AI developers ask for: digitized books, digitized newspapers, images and visual elements extracted from scans, and archival records, and each carries its own rights question. The table gives one example each, as of 2026-11-16.

Holding Example on record Rights question Source
Digitized books Institutional Books 1.0: 983,004 volumes of 1,075,899 scanned, 242B tokens (2025-06-10); 254 or 379 languages (both on the same page) copyright status per volume; release terms arxiv.org; institutional.org
Digitized newspapers Institutional Newspapers (Boston Public Library): 1.47 million pages, 16.3 billion tokens (2026-09-02) publishers' and authors' rights institutional.org
Images and visual elements from scans Institutional Books: 22.6 million visual elements extracted (2026-09-08) rights in each image; third-party works institutional.org
Archival records with personal data news archive in the GEDI and OpenAI agreement: special-category and judicial data (Garante, 2024-11-27) GDPR legal basis, notices, right to object garanteprivacy.it

The Institutional Books release states no rights breakdown. Memory institutions are one entry among every type of seller of AI training data.

Owning a book, print or record does not give an institution the copyright in it, so a collection can license only what is out of copyright or what the institution holds rights in. Three groups:

  • out of copyright: Directive 2006/116/EC (in force January 16, 2007) sets EU protection at the life of the author plus 70 years (Art. 1(1)) from January 1 of the following year (Art. 8); 17 U.S.C. 304(b) gives a pre-1978 US work in its renewal term 95 years from the date copyright was originally secured, and 17 U.S.C. 105(a) withholds copyright from any work of the United States Government
  • in copyright, with the institution as rights holder: works whose rights the institution holds
  • in copyright, with third parties as rights holders: works needing the rights holders' permission or a collective agreement

Recital 55 of Directive (EU) 2019/1024 applies the same logic: a document initially owned by a third party with unexpired protection counts as third-party IP. Term calculation is set out under public domain and openly licensed material in AI training.

Collection Licensing Routes on Record

Two routes involving memory institutions are on public record: open release with commercial use on request, and a library licensing in-copyright works for its own model training; no library, archive or museum has a published direct license to an AI developer with a price. The table compares the routes, as of 2026-11-16.

Route Institution Terms as published Money terms published Source
Direct license by a library, archive or museum to an AI developer none on public record n/a n/a n/a
Library as licensee for its own models National Library of Norway, from Kopinor (signed 2025-12-19) non-exclusive; Norwegian press publications "up to one year from the date of publication" (s.2); models may be shared publicly under license terms; rights holder opt-out for future training (s.16) NOK 45,000,000 for 2026, index-adjusted from 2027 (s.17), paid by the library kopinor.no agreement PDF
Open release, commercial use on request Institutional Data Initiative (Harvard Law School Library) noncommercial early-access terms; commercial use including AI model training requires contacting the initiative not published huggingface.co, checked 2026-11-09

Open Release With Commercial Use on Request: A Law School Library

The Institutional Data Initiative, a research initiative at Harvard Law School Library, publishes library collections as data and asks commercial users, including for AI model training, to contact it. It was born from the Library Innovation Lab.

Institutional Books 1.0 holds 983,004 of the 1,075,899 volumes scanned, with 242B tokens counted by the o200k_base tokenizer (arXiv, 2025-06-10), and was open-sourced on June 12, 2025; the initiative's page gives 254 languages in its summary and 379 detected.

The dataset card on Hugging Face applies noncommercial early-access terms (checked November 9, 2026). The initiative later added 22.6 million visual elements (September 8, 2026) and Institutional Newspapers (September 2, 2026). The dataset itself is covered under Institutional Books contents and access terms.

A National Library on the Paying Side of a Training License

The National Library of Norway licenses Norwegian press publications from Kopinor to train Norwegian and Sami language models, so here the library is the licensee and pays. Nasjonalbiblioteket is a state library whose project Mimir trains the models.

The agreement, signed on December 19, 2025, is non-exclusive and allows models to be shared publicly under license terms; the remuneration is NOK 45,000,000 for 2026, index-adjusted from 2027 (s.17).

Section 2 limits the scope to press publications published in Norway "up to one year from the date of publication", and excludes images, audio, video, specialized press, magazines and weeklies.

Norway's Copyright Act (LOV-2018-06-15-40) holds no text and data mining provision in its consolidated text (read November 9, 2026), and extended collective licensing under section 63 enabled the deal. Kopinor states that the 2027 state budget proposal has no funding for its books agreement with the library (October 26, 2026).

The payment runs from a library to a collective, and the clauses of the Kopinor and the National Library of Norway agreement are set out in full.

Charging and Exclusivity Rules for Public Collections in the EU

Publicly held collections in the EU fall under Directive (EU) 2019/1024, which governs whether libraries, museums and archives can charge for or exclusively license documents, and it excludes documents in which third parties hold intellectual property rights. It entered into force on July 16, 2019 (transposition due July 17, 2021).

Implementing Regulation (EU) 2023/138 requires high-value datasets to be free of charge, machine-readable, available through APIs and under CC0 or CC BY 4.0 or a less restrictive license, from June 9, 2024; its recital 6 exempts libraries, including university libraries, museums and archives, from the free-of-charge requirement. The directive's own articles decide what a collection may charge.

The general re-use rules of public bodies are covered under governments and open data as AI training data sources.

What Cultural Heritage Institutions May Mine Without a License

EU law lets cultural heritage institutions mine works for scientific research without a license, and that right cannot be removed by contract, but it covers the institution's research, not a developer's commercial training.

Directive (EU) 2019/790 Article 3 grants it, Article 7(1) protects it from contract override, and transposition was due by June 7, 2021. National versions include:

  • Germany: section 60d UrhG, which excludes organizations controlled by a private company with privileged access
  • Finland: 13 b § 2 mom. of Act 404/1961, which contracts or technical means cannot override
  • Hungary: section 35/A (2) of Act LXXVI of 1999 (in force June 1, 2021), with secure storage
  • Czechia: section 39d of Act 121/2000, with copies stored securely and kept for research and verification

Commercial mining by developers follows the EU text and data mining exception and the opt-out, where Article 4(3) lets rights holders reserve the right in machine-readable form for online content.

Personal Data Inside Archival Collections

Archives hold personal data, including special categories and criminal records, and a regulator has warned that licensing such an archive for AI training can breach the GDPR.

The Italian Garante's provvedimento n. 741 of November 27, 2024 warned, under Article 58(2)(a) GDPR, over the September 24, 2024 agreement between the publisher GEDI Gruppo Editoriale and OpenAI.

The Garante found special-category and judicial data in the archives, an impact assessment that did not adequately analyze the legal basis, unmet information duties and no guaranteed right to object.

A recipient of personal data informs data subjects within one month at the latest (GDPR Art. 14(3)(a)), unless that is impossible or disproportionate (Art. 14(5)(b)); legal basis and notices follow the GDPR rules on AI training data.

How AI Developers Reach Collections Without a License

AI developers can reach much of what memory institutions hold without licensing it from them: by buying and scanning print copies, crawling what is published online, or downloading open releases. Three routes:

  • buying and scanning print copies, which needs no license from an institution
  • crawling collections published online, unless reserved under Article 4(3) of Directive (EU) 2019/790
  • retrieving open releases under their terms, such as Institutional Books

In Bartz v. Anthropic (N.D. Cal. 3:24-cv-05417), the Order on Fair Use of June 23, 2025 (Dkt. 231) held digitizing purchased print books fair use and did not excuse pirated library copies; the case settled, with final approval on July 20, 2026 (Dkt. 680).

The court record on each route is covered under books as AI training data.

License Terms a Memory Institution Should Read Closely

A memory institution needs terms that track its mission: which uses are commercial, what happens to models after the license ends, and how rights holders can leave. Four terms matter:

  • commercial versus noncommercial use
  • whether trained models may be shared and survive termination: under the Kopinor agreement, termination (first effective January 1, 2028) does not affect models already trained and shared
  • the rights holder opt-out for future training (Kopinor agreement s.16)
  • excluded model uses, such as specialized models delivering media and news services

Under Regulation (EU) 2024/1689 Article 53(1)(d), applicable from August 2, 2025, the training content summary shows licensed datasets only as Yes or No plus modality. Drafting guidance is under field of use and scope limits in AI data licenses.

Licensing or Not: Benefits, Risks, Downsides and Alternatives for Memory Institutions

Licensing can fund digitization and set terms for use that would otherwise happen anyway, but it can conflict with rights holders, donors, privacy and public access.

Benefits

  • a priced agreement exists, with a library paying
  • commercial use on request keeps control over a release
  • digitized collections have AI value as released

Risks

  • archives with special-category data draw regulator warnings
  • third-party intellectual property is excluded from public re-use rules
  • public funding can lapse

Downsides

  • licensed data appears only in coarse form in training summaries
  • no published price exists for a library as seller
  • scanning purchased copies needs no license

Alternatives

  • open release under noncommercial terms
  • a collective route
  • a machine-readable reservation under Article 4(3) for online collections
  • not licensing at all

The general decision is set out under whether to sell or license your data.

How Memory Institutions Differ From Universities, Collectives and Stock Libraries

Memory institutions differ from universities in what they hold: a university licenses research data it produced, while a library, archive or museum mostly holds works created by others. Research data is covered under universities and research groups licensing data. Related types are:

The rules on this page come from statutes, a regulator's decision, a court docket and the institutions' own pages, each dated, and this page is general information and not legal advice. Further pages:

Frequently Asked Questions

Can a library license its digitized collection to AI companies?

Yes, for works out of copyright and works it holds rights in; works owned by others need the rights holders' permission.

Is Institutional Books free to use for AI training?

Not for commercial training: the dataset is offered under noncommercial early-access terms, and commercial use including AI model training requires contacting the Institutional Data Initiative. (checked November 9, 2026)

Do AI companies pay libraries for their collections?

No payment by an AI developer to a library, archive or museum is on public record; the one priced agreement has a library paying a collective. The NOK 45,000,000 for 2026 is a library's payment, not a developer's. Valuation methods are explained under how buyers value data and public price points.

Can a museum or archive license works it holds but does not own?

Only with the rights holders' permission or under a collective agreement that covers them, such as the extended effect Kopinor's agreement provides for. That effect depends on approval under section 63 of the Norwegian Copyright Act (s.3).