1. Introduction

This article examines whether Nordic collective management organisations (CMOs) license AI-related uses of copyright-protected content in a manner consistent with Directive (EU) 2019/790 (DSM Directive) and Directive 2014/26/EU (CRM Directive). The focus is not on whether the legislative framework is appropriate, but on whether current licensing practices comply with existing EU law.

While the precise scope of the text and data mining (TDM) exceptions in articles 3 and 4 of the DSM directive remains uncertain, in particular in relation to use for the training of generative AI models and the provision of subsequent AI services, some issues must be considered quite clear, such as the need for permission for including works and other subject matter protected by copyright (protected content) in the output of AI. On the other hand, it seems equally obvious that permission from rightholders is not required for uses clearly covered by copyright exceptions. The question therefore arises whether CMOs’ practice of extending licences to uses covered by copyright exceptions complies with the licensing rules of the CRM Directive, in particular the duty to conduct negotiations in good faith.

2. The text and data mining exceptions in article 3 and 4 of the DSM directive

Before examining the licensing approaches of Nordic CMOs, let us first look at the TDM exceptions in articles 3 and 4 of the DSM directive. Article 3 concerns TDM for the purpose of scientific research, whereas article 4 concerns TDM for any other purpose.

Article 3 is a mandatory exception that prevents contractual restrictions from undermining the possibility for research organisations and cultural heritage institutions to conduct TDM on content to which they have lawful access^{(1) Article 3 para 1 states that Member States shall provide for an exception for reproductions and extractions made by research organisations and cultural heritage institutions in order to carry out, for the purposes of scientific research, text and data mining of works or other subject matter to which they have lawful access. The scope of application of article 3 is broad; it concerns all different types of protected subject matter, i.e. works, fixations of performances, phonograms, films, fixations of broadcasts, databases and the new press publishers right, which was introduced through the DSM directive. Article 3 also allows the retention of the copies made for TDM purposes, as it allows copies to be stored as long as necessary and used also for verification of research results. This in fact would also seem to allow giving access to the copies made in order to verify the results of a TDM process. However, such access may not be provided to the public, but must be restricted to persons that need such access in order to verify the research. The concept of “lawful access” is also quite broad, including not only licensed content but any content freely available on the internet (with the exclusion of pirated content).}. Its background lies in conflicts between the research community and publishers regarding the use of computational tools on licensed scientific publications^{(2)Regarding the background and purpose of article 3, see recital 10 of the DSM directive.}. The purpose was not merely to create a defence against infringement claims, but to provide a genuine right to conduct TDM and prevent publishers from neutralising the exception through contractual or technical restrictions.

Although article 3 allows rightholders to apply measures necessary to protect the security and integrity of their systems, it significantly restricts their freedom to control how lawfully accessible content may be exported, analysed and retained for research purposes.

Article 4 follows a different logic. Unlike article 3, it does not prevent contractual arrangements that limit TDM. Most importantly, it allows rightholders to opt out by reserving rights expressly and in an appropriate manner. The exact requirements for such a reservation remain uncertain and will ultimately require clarification by the European Court of Justice (ECJ).

The question therefore arises whether CMOs’ practice of extending licences to uses covered by copyright exceptions complies with the licensing rules of the CRM Directive, in particular the duty to conduct negotiations in good faith.

The distinction between Articles 3 and 4 suggests that Article 4 leaves considerably more room for service providers to structure their business models, including the deployment of proprietary TDM tools. Copies made under Article 4 may only be retained for as long as necessary for mining purposes. This prevents the indefinite retention and reuse of protected content. The scope of article 3 is also limited to a very special use case, that of scientific research. Taken together, these differences reflect the application of the three-step test.^{(3)DSM directive article 7.2.}

Article 4 was introduced to avoid the possibility of an e contrario interpretation of article 3 that reproduction of protected content necessary to perform TDM for other purposes than for research would require permission from the rightholders.

Contrary to the general approach to exceptions and limitations to copyright in the Infosoc-directive, articles 3 and 4 of the DSM directive are mandatory exceptions that as a starting point provide little room of manoeuvre for Member States.

There are several safeguards in the DSM directive which are aimed at preserving freedom of information. Firstly, it is clarified in the definition of TDM that information retrieval or analysis of works are not in itself copyright relevant, but the exceptions and limitations to copyright in article 3 and 4 only cover reproductions and extractions made with the purpose of making it possible to analyse protected content through computational means. Secondly, regarding the new press publishers right, it is also clarified that copyright does not cover mere facts and that the new right is subject to the exceptions and limitations of the Infosoc directive, including the exception in the case of quotations for purposes such as criticism or review. This also has an impact on the interpretation of the TDM exceptions.

Thirdly, although search engines are not mentioned expressly, it is apparent that one of the main aims of article 4 is to enable search engines to continue to operate, as they are essential to the functioning of the internet. In practice they should be protected from claims of infringement of copyright as long as they fulfill the requirements of the directive, in particular do not index webpages excluded from indexing such as private areas (intranet or other areas with restricted access) and honor robot.txt and similar technical measures.

3. The room for manoeuvre left to Member States in relation to articles 3 and 4

There is little indication that Member States enjoy substantial discretion when implementing Articles 3 and 4. The DSM Directive requires Member States to introduce these exceptions. Although Article 4 is formally labelled an ‘exception or limitation’, this does not imply freedom to fundamentally alter its operation. Its core structure consists of a TDM exception combined with an opt-out for rightholders^{(4)Because of the opt-out-possibility, it is a merely a limitation.}.

Recital 17 of the DSM directive states that Member States are not to provide for compensation for rightholders as regards uses under the text and data mining exceptions (in plurals) introduced by the Directive. This also suggests that Member States are expected to transpose the articles without modifying their modus operandi.

4. The application of the TDM exceptions to AI

The DSM Directive defines TDM as an automated analytical technique aimed at analysing digital text and data to generate information, including patterns, trends and correlations. TDM itself is not a copyright-relevant activity. Copyright becomes relevant only because reproductions of protected content may be required.

The TDM exceptions express the principle “a right to read is a right to mine”. The basic idea underlying Articles 3 and 4 is that reproductions made solely to analyse protected content should not affect its commercial exploitation.

As explained in recital 9 of the directive, TDM can also be carried out in relation to mere facts or data that are not protected by copyright, and in such instances no authorisation is required under copyright law. There can also be instances of TDM that do not involve acts of reproduction or where the reproductions made fall under the mandatory exception for temporary acts of reproduction provided for in the Infosoc-directive, which should continue to apply to TDM techniques that do not involve the making of copies beyond the scope of that exception.

Text and data mining refers to an activity and shall apply notwithstanding the type of TDM tool or methodology applied. Examples of different TDM methodologies are statistical tools, search engines, machine-learning models and foundation models.

A key question is whether the TDM exceptions apply when analysing protected content with the help of AI, when training foundation models or AI systems and when creating outputs of AI.

It seems quite clear that articles 3 and 4 at least cover the reproduction necessary for providing the input to be analysed in an AI system. It is also very clear that the exceptions in articles 3 and 4 do not cover the output. This means that the reproduction, communication or making available to the public of output that includes protected content requires the authorization of the rightsholders.

The position regarding AI training is more complex. Training foundation models appears to satisfy the definition of TDM because it involves automated analysis of text and data to generate information and capabilities. Recital 18 further acknowledges that TDM supports the development of new applications and technologies. At the same time, copyright concerns may arise where protected content remains incorporated in ways that extend beyond what is necessary for analysis. Whether particular training practices fall inside or outside the TDM framework must therefore be assessed case by case and may ultimately require judicial clarification by the ECJ.

5. Licensing practices of Nordic copyright societies

Nordic CMOs have adopted different approaches to AI-related licensing. Some rely on ordinary collective licensing while others employ extended collective licensing (ECL) mechanisms. The scope of the licences also varies considerably. Some focus on AI training, others on prompting, and some potentially extend to outputs.

The Swedish CMO for musical works STIM has introduced a licence covering rights required for certain AI models and services, although the repertoire covered remains limited^{(5)https://www.stim.se/nyheter/stim-lanserar-varldens-forsta-ai-licens-for-musik.} . In Finland, Teosto has not yet introduced comparable licensing but has amended its authorisation framework in preparation for AI-related licensing^{(6)https://www.teosto.fi/tekoaly-oikeusluokka/.}. It is unclear whether their licensing conditions would also cover outputs generated by AI, although this possibility cannot be entirely excluded.

Kopinor in Norway has concluded an agreement with the National Library concerning the use of news content for training Norwegian and Sami language models^{(7)https://kopinor.no/aktuelt/historisk-avtale-om-avisinnhold-til-kunstig-intelligens/. The conditions of the agreement is found at https://kopinor.no/wp-content/uploads/2026/02/Trening-av-norske-sprakmodeller.pdf.} . The agreement is designed to facilitate public-interest language technology while excluding specialised services that would compete with media or news publishers. The agreement appears primarily concerned with training, and it is not entirely clear whether it also covers AI outputs to the extent that protected content may incidentally be included in the output.

Kopinor has adopted a cautious approach to licensing reproductions for educational purposes. Its agreement with universities and higher education institutions excludes machine-learning and generative-AI training unless supported by a separate legal basis^{(8)This structure seems to indicate, that the agreement does not preclude the possibility for the TDM exceptions to be applied.}. A forthcoming private-school agreement permits limited prompting of AI systems in closed environments while excluding systematic use for training internal or external AI solutions^{(9)§ 8 of the agreement, https://kopinor.no/wp-content/uploads/2026/06/Privatskoler-2027.pdf.}. Similarly, the Bokhyllan agreement excludes the use of digitised cultural heritage collections as a basis for AI training without a specific licence and prohibits indexing by external search-engines.

In Denmark Tekst & Node has received approval from the Danish Ministry of Culture to provide an addition for prompting purposes to existing extended collective licences for reproduction of reprographic works^{(10)https://kum.dk/fileadmin/_kum/2_Kulturomraader/Ophavsret/Aftalelicensgodkendelser/Generelle/Tekst___Node_-_OHL____50__stk._4__jf.____50__stk_2.pdf}. The licences cover educational establishments and internal use in companies and other organisations. The licence conditions have not been published.

In Finland Kopiosto has added to its licences for reproduction or use for internal purposes in private sector organisations^{(11)https://kopiosto.fi/lupakauppa/wp-content/uploads/sites/3/2025/10/Kopiointiluvan-lupaehdot-yritykset-ja-muut-yhteisot.pdf}as well as for municipalities and well-fare areas^{(12)https://kopiosto.fi/wp-content/uploads/2022/06/Kopiointiluvan-lupaehdot_kunnat-ja-hyvinvointialueet.pdf}a right to use works (text, pictures) for prompting in internal, secure AI systems. The licence does not allow for the use of generated content for external use or for prompting in other than internal, secure AI systems and does not allow, for instance, pictures to be modified with the help of AI.

Kopiosto offers in addition an add-on licence that allows for finetuning of AI systems for internal use^{(13)https://kopiosto.fi/lupakauppa/wp-content/uploads/sites/3/2025/10/Sisaisen-tekoalyohjelman-hienosaatamisen-lupaehdot-yritykset-ja-muut-yhteisot.pdf}. The licence is confined to protected content to which the organisation has legal access^{(14)According to the agreement, “legal access” is based on subscription, agreement or purchase. The concept of legal access is thus more limited than the definition in the DSM directive, which specifically covers any protected content made available on the internet with the rightholders permission.}and provided that the finetuned AI system is only used within the organisation. The licence excludes a wide range of protected content such as the collections of cultural heritage institutions, research reports and similar publications and content the use of which is based on a specific agreement or licence. Specifically excluded is the possibility to collect or record works with web scraping or corresponding data collection methods or use data sets collected by a third party for fine tuning. Specifically excluded is also the possibility to use protected content in whole or in part for TDM^{(15)It seems contradictory to prohibit TDM while licensing AI training. The relationship between these concepts have not been explained in conjunction with the agreement.}.

As these examples demonstrate, there are substantial differences in licensing strategies of CMOs. Unfortunately many CMOs don’t offer public access to the actual texts of the agreements for an in-depth analysis. Nevertheless, it seems plausible that in most (if not all) cases they do not license the use of protected content for training of general AI models or publicly accessible AI services. Such licensing can instead be expected to emerge predominantly through direct contractual arrangements^{(16)In the music sector Spotify is said to negotiate a licensing agreement related to AI with the majors, https://www.bbc.com/news/articles/cn7emv83edjo.} .

6. Assessment under EU law

When assessing whether some of the agreements offered by the CMOs could be unlawful, an important question is whether a use is covered by an exception and whether that exception is mandatory.

To the extent agreements made by CMOs cover the output of AI, there do not seem to be any particular legal challenges, providing that the authorizations from the rightholders or the ECL arrangement of the country allows for such licences. Such use is not part of the TDM exception.

There might also be agreements that cover uses that are in the grey zone. Even though training of AI appears to fulfill the general requisites and the definition of TDM in the directive, there might be other reasons why the exceptions may not be applied (for instance if a copy is retained for an unlimited period of time).

Making copies necessary to analyse protected content (for example through prompting) seems to be part of the key area of the mandatory exceptions in article 3 and 4; that activity can clearly be considered to fulfill the principle of “a right to read is a right to mine”. But even in that case, the exception in article 4 is not applied if copyright has been reserved expressly and in an appropriate manner^{(17)Contrary to this, in relation to article 3 it is clear that no room for licensing solutions exist under that TDM exception.}.

Could a CMO be considered to have the capacity to represent rightholders in expressing an opt-out in accordance with article 4? In principle I don’t see why not, provided that the CMO has been duly mandated to do so and possesses the practical means to implement it.

However, practical difficulties arise insofar as CMOs don’t maintain comprehensive databases of protected content and lack the technical mechanisms that would make work-specific opt-outs readily identifiable. The structure of Article 4 instead seems to assume that the commercial actor placing content on the market is typically best positioned to reserve rights.

Another question is whether Member States may effectively replace TDM exceptions with extended collective licensing arrangements. Article 12 in the DSM directive recognises licensing mechanisms with extended effect but also states that it does not affect other provisions of Union law, including provisions containing exceptions and limitations. In my view, this cannot be interpreted as allowing Member States to transform mandatory exceptions into permission-based licensing schemes.

7. The CRM directive requires CMO’s to act responsibly in their licensing activities

The CRM Directive recognises the gatekeeper role of CMOs and therefore requires licensing activities to be conducted on objective, reasonable and non-discriminatory terms. Negotiations must be carried out in good faith and tariffs must reflect the economic value of both the rights and the licensing services provided.

Against this background, concerns arise where CMOs attempt to license activities already covered by mandatory TDM exceptions. Acting in good faith arguably requires informing users when a particular use does not require authorisation. Efforts to present statutory exceptions as licensable rights sit uneasily with the objectives of the CRM Directive.

Particular concerns arise in relation to some prompting licences, where users may effectively be paying for uses that are exempted from copyright. As a minimum CMOs would be wise to cautiously draft the agreements to only complement existing exceptions, thus ensuring that the agreements do not contravene the possibilities to use a work in accordance with a copyright exception.

The current practice may also create difficulties for rightholders. If broad rights-reservation statements by CMOs are assumed to constitute valid opt-outs under Article 4, rightholders wishing to permit TDM may have to withdraw from collective arrangements in order to ensure that their works remain available for TDM. Such an outcome would run counter to the original logic of Article 4.

However, it remains to be seen whether users, rightholders or supervising authorities will intervene in the licensing practices of CMOs, to the extent made possible in the national laws transposing the CRM directive, or if the CMOs themselves correct their current practices to reduce legal risks.