AI and copyright: is a licensing market just a question of time?
What transpires from recent news about Meta’s use of a pirate website to train its AI model is certainly not good news for authors. Indeed, very disappointed but not surprised, the authors have entered 2025 with a further confirmation that protected works have been and will continue to be used without their authorisation. The news arrives at a crucial and turbulent time for the EU and its authors, for at least two reasons: the battle on interpretation of the text-and-data mining exception and the drafting process of the AI Code of Practice.
Firstly, Regulation 2024/1689 (AI Act) enshrines Art 4 Directive 2019/790 (hereinafter: DSM Directive), extending the text-and-data mining (TDM) exception for commercial purposes to AI. In practice, a copyright exception created with a certain technology (TDM) in mind, has been extended to a technology (AI) that in 2019 was far from being the generative tool that the public is now fond of using. As a consequence, billions of works are being scraped from the internet and generating billions in turnover for AI companies, with no prior authorisation from the authors and, consequently, with no remuneration. The only way authors are supposed to be able to escape from this is by declaring they are opting-out… but more on this later.
Secondly, at the time of writing, the second draft of the AI Code of Practice has just been discussed by stakeholders in the Transparency and Copyright Working Group, of which the SAA is a participant. As some may know, Art 56 of the AI Act provides that the AI Office should ‘facilitate the drawing-up of codes of practice’ which are specifically aimed at implementing the AI Act in its entirety. This drawing-up process consists of periodical meetings of four working groups created according to different themes, including copyright. Participants have limited time to send comments to each draft and some of them can express their views in three minutes at each meeting, with no further discussion. The full process can be found here.
In spite of the AI Code of Practice drafting process keeping us busy, the AI Act is merely a small part of the story and only the beginning. The uncertainties around the AI and copyright story are not only affecting the EU and news such as those mentioned above are equally important for authors worldwide. Notably, one can mention the US and its more than 30 pending lawsuits against AI companies, and the UK where a consultation has been opened in December 2024 and is set to close in late February 2025.
However, on the brighter side, these uncertain times also seem to attract discussions at the EU Council, in its Copyright Working Party, first with the Hungarian EU Presidency and more recently with the Polish EU Presidency. The two presidencies’ initiatives aim at taking the reflection around AI and copyright beyond the AI Act by focusing on potential solutions that could ensure the remuneration of authors, including with the support of a collective management architecture.
The Hungarian EU presidency asks, Member States reply on AI and copyright
In June 2024, the Hungarian Presidency published a questionnaire titled ‘Policy questionnaire on the relationship between generative Artificial Intelligence and copyright and related rights’. The questionnaire was only open to EU Member States to reply and was discussed at the meetings of the Council Working Party on Copyright. The questionnaire addressed seven themes: state of play in the field of AI and copyright; training of AI models; the protection of outputs; transparency; remuneration; liability for copyright infringement; and policy and international context. The SAA provided comments to its members (which they could share with their Ministries) and also sent a contribution to the Hungarian Presidency.
The Hungarian Presidency prepared a summary of the Member States’ responses. The summary allowed to observe EU countries’ views on the matter and provided a bit of hope and for a moment relieved the tension of the ‘opt-out or nothing’ reality that had been forced ever since the inclusion of the TDM exception in the AI Act (and the opt-out’s failure of being a mechanism to facilitate the conclusion of licensing agreements between rightsholders and AI companies).
With regards to the training of AI models, the Hungarian summary reports (p. 10) that: “According to a significant number of contributions, the applicability of the TDM exception to AI training is unclear” and “a few Member States […] added that the TDM exception – if applied in a broad sense, covering all types of AI training – would not respect the three step test principle”.
These types of assertions are very welcome by the SAA who in its position paper on AI published in 2023, warned against the application of the TDM exception to generative AI and the conflicting effects it would have with the internationally-recognised three-step test. Indeed, the three-step test mandates that exceptions to copyright shall be specific, should not impede the normal exploitation of protected works and should not interfere with the legitimate interest of the authors (and rightsholders). Extending an exception from TDM to AI training, already breaks the grounds of specificity upon which the three-step test stands. Moreover, as AI is trained to create works that compete and affect the value and market of human works, the rest of the test even more decisively crumbles in front of the TDM exception applied to AI training.
This standpoint is reinforced by a much discussed article by Tim W. Dornis, according to whom there is a significant difference on a technical level between TDM and AI, hence making the TDM exception not applicable to AI training.
On a similar note, the remuneration section of the summary highlights the importance of collective management organisations (CMOs) in ensuring that authors are properly remunerated in the context of AI uses (p. 21 and 23): “A significant number of Member States were of the view that the mass scale of AI-related copyright uses and the practical challenges of the rightholders in monitoring them both underline the importance of collective management in this area” and “a remuneration scheme shall be guaranteed”. The Member States have shared their belief that a system of mandatory collective management or extended collective licensing can streamline licensing and remuneration in the context of AI.
The SAA has highlighted this aspect on several occasions, including in its own reply to the Hungarian questionnaire and in the drafting process of the AI Code of Practice. It is almost a given that in such contexts, wherein one can find mass use of protected works, CMOs are best positioned to ensure that authors are properly remunerated. This idea is true for at least two main reasons: the efficiency and collective power of CMOs.
I. Efficiency: the amount of works used in the development of AI would make other non-mediated solutions, such as direct licensing, more difficult. As CMOs have a mandate for vast repertoire of works, their activity is pivotal to ensure that all works are covered and all remunerations collected are redistributed properly. This is clearly a more efficient (and effective) solution than an alternative reality wherein the authors must knock at every AI company’s door.
II. Collective power: in the absence of CMOs, the single authors should go to each and every AI company to negotiate licensing agreements for each and every work they have created. On the contrary, CMOs have a long-standing position in the market of copyrighted works and have ensured remuneration for the authors they represent even with worldwide powerful users. As AI companies are often backed by Big Tech such as Google, Meta and Microsoft, this aspect only reinforces the necessity to accept collective licensing as a fair solution.
Last but not least, in the section of the questionnaire dedicated to the liability for copyright infringements, the Hungarian Presidency reported that some Member States have already thought about their plans concerning the liability of AI companies, both in the input (i.e., when the AI is fed data to learn) and output (i.e., when AI generates something upon request of the user) phases of AI. In their response, although the Member States specified it might be too early to architect a liability regime, Art 17 DSM Directive has been mentioned as a source of inspiration. In said article, intermediaries take the responsibility for infringing behaviour, releasing the end-users from a liability that would be difficult – if not impossible – to enforce. In the vision of the respondents (p. 28), regardless of the type of liability chosen, it “should cover both the input and output related infringements”.
The SAA shared similar views on the matter. In the comments provided to this part of the questionnaire, it was important for the SAA to highlight that Art 17 DSM Directive sets an obligation for the intermediaries to obtain an authorisation for the works made available to the public by their users. In the context of AI, providers must demonstrate they have made their best efforts to obtain an authorisation, both for the input and for the output. This way, the SAA envisages the creation of a healthy and fair licensing market for AI uses which, at this point in time, is non-existent. The same healthy and fair licensing market that is hoped for by the Polish Presidency in a new questionnaire.
The Polish Presidency asks, Member States answer on CMOs challenges
On 15 January 2025, the Polish Presidency published a questionnaire titled ‘Policy questionnaire on the Challenges facing Collective Management Organizations in the EU Member States’. Seemingly inspired by its predecessor, the Polish presidency wanted to continue the dialogue on AI opened by the Hungarian Presidency, now however more focused on the challenges that CMOs are facing due to the unstoppable advancement and spread of technologies.
The new questionnaire proves that some of the opinions shared during the Hungarian Presidency have not remained unheard: there is a recognition of the important role CMOs have in these matters to ensure a healthy development of AI technology in the EU and beyond. In contrast to the Hungarian approach, Member States are further asked to seek the assistance of CMOs when answering the questions.
The Polish questionnaire inquires on topics ranging from the (i) challenges AI brings to CMOs to the (ii) relationship between CMOs and online platforms and the (iii) transparency of CMOs. In the case of (i) and (ii), important issues for the interest of audiovisual authors and their works in the digital realm – especially with the hasty development of AI – are raised.
In particular, pertaining to the challenges brought to CMOs by AI (i), the questionnaire asks Member States what measures could assist CMOs to conclude agreements for AI uses, and what measures are currently being taken by CMOs to tackle infringement in the context of AI use. With regards to the relationship between CMOs and online platforms (ii), the questionnaire inquires on whether the Member States have adopted or have the intention to adopt measures that would ensure remuneration for authors in the evolving technological environment, within the framework of Art 18 DSM Directive (introducing an EU-wide principle of appropriate and proportionate remuneration for authors).
The perspective adopted by the Polish Presidency seems to hit the problem at its very core and promises to incentivise meaningful discussion for the future of authors’ creations in the age of AI, wherein CMOs have a central role in ensuring the respect of the remuneration principles set in the EU copyright acquis, now in the AI context.
As hinted earlier, following the entry into force of the AI Act and the inclusion of Art 4 DSM Directive within the realm of the AI regulation, the only thing in the hands of authors has been that of opting out, meaning that the authors communicate that their works should not be used for AI purposes. This rights reservation option – in theory – means that the authors’ works which have been opted out should be erased from AI models that have been on the market ever since before the entry into force of the AI Act and should not be used for developing new AI models. As a consequence, AI companies should be encouraged to enter into licensing agreements with authors who have opted-out. However, except for very rare cases, none of this is really happening.
In fact, there is no current understanding on how it is feasible, in technological terms, to ‘untrain’ the AI (or, if there is a way, it is kept hidden or considered too costly). Therefore, the works that have been used to develop the AI models that are available on the market are already being infringed. If this was not enough, AI companies do not seem to care about licensing and new AI models are being developed without proper copyright compliance. Or rather: they do not care unless it is their own work that is being used with no authorisation, as the comments by OpenAI to the new AI chatbot DeepSeek suggest.
On top of all this, AI companies defend the exploitation of opted-out works by using the shield of the ‘machine-readable’ requirement. According to Art 4 DSM Directive, the rights reservation must be expressed in machine-readable means, which limits the ways the author can express the choice of opting-out. In turn, purely technological practices such as robots.txt are being paraded as standards for opting out, despite of their limits and without much afterthought.
In particular, these machine-readable means should respect state-of-the-art technologies, as specified by Art 53(1)(c) AI Act. According to a recent decision of the Hamburg Regional Court (Germany), this would mean that modern AI technology should already be able to read natural language – i.e., it should already be able to read a rights reservation declaration written by any of the authors or CMOs they are represented by.
As a result of the current state of uncertainty, the situation is far from encouraging the development of a licensing market and is instead entirely on the side of AI development that is blind to respecting basic copyright protection - and is even blinder if one considers that authors should be appropriately and proportionately remunerated, as Art 18 DSM Directive provides.
It’s a question of time
While options other than the application of the opt-out have until now been dismissed, the issues raised by the Polish Presidency questionnaire represent an important venue for CMOs to fully share their concerns about the current situation and to express their views.
It is impossible not to value the importance of human creativity in the AI context. AI training and development require high amount of high-quality data to be ‘ingested’. It is widely known and accepted that for AI to properly develop, high-quality data is needed. High-quality data means works created by humans, including copyrighted works. These high-quality data are sought after in opposition to synthetic data, which stand instead lower in quality as they are already a by-product of human-created works. Synthetic data has been proved to make the AI less intelligent and more prone to hallucinations (i.e., AI will start to lie), as also reminded by a report commissioned by the French Ministry of Culture.
In a context where the future of AI technology depends on the synergic collaboration of authors, it is clear that authorisation and remuneration for the use of their works is due. Taking stock of this, regardless of there being a technology that is able to respect the opting-out or not, and regardless of there being a technology that is even able to ‘read’ the opt-out declaration, creating a whole industry on the shoulders of authors without there being any form of remuneration for them will only lead to a lose-lose situation that would lead to a direct opposition between the supply and the demand, between authors and technology innovators. If authors would be properly remunerated, they will be more willing to participate in and contribute to the ‘AI revolution’.
If the author has no moral opposition against AI, permission to use their works merely depends on finding an agreement with AI companies. More practically and given the vast amount of content used to train AI, CMOs are best placed to ensure a smooth flow of remuneration from AI companies to authors, for the creation of a healthy and fair licensing system. This solution cannot be avoided for too long and it might be just a question of time, for the benefit of both authors and the future of AI.
SAA Legal and Policy Officer