legal247

Text and data mining in Norway: a new exception, and what are AI vendor indemnities worth?

In brief

The Norwegian Copyright Act currently has no text and data mining exception. Prop. 41 LS (2025–2026) proposes a new section 50 e permitting text and data mining of lawfully accessible works unless the rightholder has reserved its rights in an appropriate manner. Foreign case law is divided, and AI vendor indemnities carry broad exclusions and Irish governing law.

The Norwegian Government submitted Prop. 41 LS (2025–2026) to the Storting on 27 March 2026, proposing rules on text and data mining in the Copyright Act. The bill implements the Digital Single Market Copyright Directive (EU) 2019/790, which the EEA Joint Committee incorporated into the EEA Agreement by Decision No. 333/2023 of 8 December 2023. According to the Storting’s case page, the committee must report by 10 November 2026, and the plenary debate is scheduled for 19 November 2026. For businesses that develop or procure AI services, the bill raises questions about what the exception will permit, how foreign courts have assessed AI training, and what vendor indemnities against copyright claims are worth under Norwegian law. Ownership of what AI models produce is addressed in the article on copyright and AI.

What is text and data mining under the DSM Directive?

Under DSM Directive Art. 2(2), text and data mining is any automated analytical technique aimed at analysing text and data in digital form in order to generate information. The definition covers the copies made when an AI model is trained. Under DSM Directive Art. 3, research organisations and cultural heritage institutions may make reproductions for scientific research, and under DSM Directive Art. 7(1) that exception cannot be overridden by contract. DSM Directive Art. 4 adds a general exception for lawfully accessible works, which applies for any purpose, including commercial ones.

The general exception applies only as long as the rightholder has not expressly reserved its rights. Under DSM Directive Art. 4(3), the reservation must be made in an appropriate manner, such as machine-readable means for content made publicly available online.

What is the position in Norway before the amendments?

The Copyright Act currently contains no provision on text and data mining. The exception for temporary copies in Copyright Act § 4 covers only transient or incidental copies with no independent economic significance, and it does not apply to databases or computer programs. In our assessment, it can therefore only exceptionally cover a training dataset, which is normally retained over time and has considerable economic value. Until the new rules enter into force, copying for AI training in Norway must as a rule be based on the rightholder’s consent.

What does the Government propose on text and data mining?

The bill follows the Directive closely, placing the definition in a new Copyright Act § 50 d, the general exception in a new Copyright Act § 50 e and the research exception in a new Copyright Act § 50 f. Under the proposed section 50 e, anyone with access to a lawfully accessible work may make copies of it for text and data mining purposes. The copies may be retained for as long as necessary but not used for any other purpose. Under the third paragraph, the permission falls away if the author has reserved the use in an appropriate manner.

The Ministry has decided not to prescribe specific methods of reservation. For online material, it considers machine-readable methods probably the only appropriate option and mentions robots.txt, HTML meta tags, HTTP robot directives, licences and technological protection measures. In its view, a general statement in natural language currently offers limited certainty, a question ultimately for the Court of Justice. The Act will enter into force on a date set by the King in Council, and both the exception and the right of reservation will then apply to content already published.

How does the exception relate to the AI Act?

Under AI Act Art. 53(1)(c), providers of general-purpose AI models must have a policy to comply with EU copyright law, including reservations under Art. 4(3). Under point (d), they must publish a summary of their training content. In the Copyright Chapter of the GPAI Code of Practice, signatories commit not to circumvent effective technological measures, to exclude websites recognised as infringing on a commercial scale, and to follow robots.txt under IETF RFC 9309 and other appropriate machine-readable protocols. The AI Act is not yet part of the EEA Agreement, but it applies to models offered in the EU, as explained in the article on general-purpose AI models.

What have foreign courts decided?

Case law on AI training is divided, and none of the decisions binds the Norwegian courts. The table covers only decisions checked against the courts’ own publications.

Case Court and date Outcome Status
Kneschke v LAION LG Hamburg 27.9.2024 (310 O 227/23), OLG Hamburg 10.12.2025 (5 U 104/24) Covered by the research exception and, according to the OLG, by the general exception because the agency’s reservation was not machine-readable Further appeal permitted
GEMA v OpenAI LG Munich I 11.11.2025 (42 O 14139/24) Lyrics memorised in the model were reproductions outside the text and data mining exception Not final
Getty Images v Stability AI High Court 4.11.2025, [2025] EWHC 2863 (Ch) A model that does not store or reproduce the works is not an infringing copy Training claim abandoned, training took place abroad
Thomson Reuters v. Ross Third Circuit 29.9.2026 Training a competing legal search tool was not fair use Petition for rehearing en banc reportedly filed
Bartz v. Anthropic N.D. Cal. 23.6.2025 Training was fair use, a permanent library of pirated copies was not Settlement approved 20.7.2026
Kadrey v. Meta N.D. Cal. 25.6.2025 Fair use because the plaintiffs failed to show market harm Expressly of limited consequence

The German judgments are the most relevant for Norwegian law, because they interpret national rules implementing the same Directive. The Hanseatic Higher Regional Court attached decisive weight to the reservation not being machine-readable. The Munich Regional Court I accepted that the exception may cover preparatory copies, but not works becoming reproducible from the model’s parameters. That reasoning bears on the proposed requirement that copies may not be used for other purposes. The American decisions rest on fair use, which Norwegian law does not recognise, but they shape the litigation risk vendors bear and thus the indemnities they offer.

What do AI vendor indemnities cover?

The major vendors undertake to defend the customer and pay judgments or settlements where a third party claims that use of the service or its output infringes intellectual property rights. The undertakings are conditional, and the exclusions are similar.

Vendor Coverage Typical exclusions Cap and governing law
Microsoft (Customer Copyright Commitment) Output from covered products Filters disabled, customer knew or should have known, no right to input, trade marks, required mitigations missing Not stated, depends on the licensing agreement
Google Cloud (Section 20(i)) Unmodified output from listed paid services, and training data Customer knew or should have known, citations or filters disregarded, use after notice, trade marks Not stated, depends on the customer agreement
OpenAI (Services Agreement and Service Terms) The services, and API and Enterprise output Customer knew or should have known, safety features unused, output modified or combined, trade marks Uncapped, Irish law and Dublin courts for EEA customers
Anthropic (Commercial Terms) Paid use, training data and output Modifications, combination, customer inputs, customer knew or should have known, trade marks Uncapped, Irish law and courts for EEA customers

For Azure OpenAI, Microsoft also requires a system message against infringement, testing for reproduction of third-party content and protected material filters, according to its required mitigations. OpenAI’s Service Terms, Section 1(a) also exclude output combined with products or services not provided by OpenAI. Under Section K.5 of Anthropic’s terms, the indemnity is the sole remedy for third-party claims.

What are the indemnities worth under Norwegian law?

Their value depends first on the governing law. For EEA customers, Section 16.3 of OpenAI’s Services Agreement and Section M.7 of Anthropic’s terms provide for Irish law and Irish courts, and such a clause will as a rule bind a Norwegian customer under Lugano Convention Art. 23. The Norwegian rule that limitations of liability do not protect against gross negligence, and adjustment of unreasonable terms under Contracts Act § 36, therefore cannot simply be invoked against the vendor. The rightholder’s claim against the Norwegian customer, by contrast, is decided under Norwegian copyright law as the law of the country for which protection is claimed.

The coverage is narrower than the customer’s exposure. Under Copyright Act § 81, even a person acting in good faith may have to pay a reasonable fee, and under Copyright Act § 82 the court may order measures against infringing products. The indemnity covers judgments, settlements and the defence, but not the cost of ceasing use, rebuilding the service or changing model. Under Section 13.3 of its agreement, OpenAI may instead terminate and refund prepaid fees. The interplay with general liability caps is discussed in the article on limitation of liability in SaaS agreements.

The exclusions also catch typical uses, since most businesses combine output with their own content and edit the result before publication. The exclusion for what the customer should have known will bite in particular once a rightholder has given notice of a claim. In our assessment, the indemnities are of real value for the core risk, namely that the model was trained unlawfully or that unmodified output reproduces protected content, but offer little protection for solutions built with fine-tuning, own data and own filters. If the vendor becomes insolvent, the indemnity is an unsecured claim, as discussed in the article on training data in insolvency. Vendor terms may also grant the vendor a right to mine the customer’s own content, as the dispute over Meta’s new terms illustrates.

What should businesses do?

Publishers should have their reservation in place by commencement, bearing in mind the separate right of press publishers explained in the article on the press publishers’ right and AI. Further material is available on the artificial intelligence topic page and the contracts topic page.

  1. Implement a machine-readable reservation in robots.txt and metadata, and add an express reservation to the terms of use.
  2. Document when the reservation was made and retain server logs showing which crawlers have visited the website.
  3. Check that the service is covered by the vendor’s indemnity, that it is paid for, and that the required filters and system messages are enabled.
  4. Negotiate away or narrow the exclusion for combination with the customer’s own content, and require cover for the cost of changing model.
  5. Establish a procedure for ceasing use of disputed output as soon as a rightholder’s claim is received, and notify the vendor within the contractual deadline.

Questions and answers

Is a reservation in a website's terms of use enough to stop text and data mining?

Not necessarily. Under the bill, the reservation must be made in an appropriate manner. The Ministry considers that machine-readable methods such as robots.txt and meta tags are probably the only appropriate method for much online content, and that a statement in natural language currently offers limited certainty. The Hanseatic Higher Regional Court took the same view when it held that a reservation which was not machine-readable did not prevent the collection.

Will the new exception apply to content that has already been published?

Yes. The Ministry states that both the exception and the right of reservation will apply from the date the amendments enter into force, including to content already available online. Rightholders who wish to opt out should therefore have machine-readable reservations in place before commencement.

Does a vendor indemnity cover an order to stop using an AI service?

As a rule, no. The obligations concern third-party claims and cover judgments, settlements and the costs of the defence. The cost of ceasing use, rebuilding a service or changing model remains with the customer unless the contract provides otherwise. Under section 13.3 of its agreement, OpenAI may instead terminate the agreement and refund prepaid fees if the service cannot be made non-infringing.

Sources
  1. Digital Single Market Copyright Directive (EU) 2019/790 Arts. 2(2), 3, 4 and 7(1)
  2. Norwegian Copyright Act (åndsverkloven, in Norwegian) Sections 4, 26, 81 and 82 (last amended by Act of 13 December 2024 No. 76)
  3. Prop. 41 LS (2025–2026), amendments to the Norwegian Copyright Act implementing the DSM Directive (in Norwegian) Section 3 and the comments on reservation and storage, proposed sections 50 d, 50 e and 50 f
  4. Ministry of Culture and Equality, consultation paper on amendments to the Copyright Act (22 November 2023, in Norwegian)
  5. Storting, case page for Prop. 41 LS (2025–2026) (in Norwegian) committee deadline 10 November 2026, plenary debate scheduled for 19 November 2026
  6. Artificial Intelligence Act, Regulation (EU) 2024/1689 Art. 53(1)(c) and (d)
  7. European Commission, General-Purpose AI Code of Practice, Copyright Chapter (10 July 2025) Measures 1.1 to 1.5
  8. Hanseatic Higher Regional Court, press release on judgment of 10 December 2025, 5 U 104/24 (Kneschke v LAION, in German) appeal against LG Hamburg, 27 September 2024, 310 O 227/23
  9. Munich Regional Court I, press release on judgment of 11 November 2025, 42 O 14139/24 (GEMA v OpenAI, in German)
  10. Getty Images v Stability AI [2025] EWHC 2863 (Ch) (4 November 2025) paras. 757 and 758
  11. Thomson Reuters v. Ross Intelligence, No. 25-2153 (3d Cir. 29 September 2026)
  12. Bartz v. Anthropic PBC, No. C 24-05417 WHA (N.D. Cal.), Order on Fair Use (23 June 2025)
  13. Bartz v. Anthropic, official settlement website (final approval 20 July 2026)
  14. Kadrey v. Meta Platforms, No. 23-cv-03417-VC (N.D. Cal.), Document 598 (25 June 2025)
  15. Microsoft Product Terms, Microsoft Generative AI Services, Customer Copyright Commitment
  16. Microsoft Learn, Customer Copyright Commitment Required Mitigations
  17. Google Cloud, Service Specific Terms Section 20(i) Generative AI Services
  18. OpenAI Services Agreement (effective 1 January 2026) Sections 13.1, 13.3, 14.2 and 16.3
  19. OpenAI Service Terms (updated 16 October 2026) Sections 1(a) and 3(b)
  20. Anthropic, Commercial Terms of Service (effective 17 June 2025) Sections K.1, K.3, K.5, L.3 and M.7

Next legal review: 30 November 2026