Why the EU TDM Exceptions (Art 3 & 4) Can Be Used for AI – Including Generative AI

3 June 2025

Blog post by Benjamin White

This blog – part of a series on recent developments around AI – explores rightsholder claims around Articles 3 and 4 of the Copyright in the Digital Single Market Directive.

As already reported by KR21, researchers and librarians have noticed a clear trend for some rightsholders to deny the applicability of the so-called “TDM exceptions” to a wide variety of AI technologies. 

In doing so, they imply that Articles 3 and 4 of the Copyright in the Digital Single Market Directive (CDSMD) (as well as Section 29A of the UK’s Copyright, Designs and Patents Act), were not devised with “artificial intelligence” in mind, but for a narrower subset of machine learning models designed for only “text and data mining.”

Such an argument is however erroneous, as can be shown both by a look at the formulation of the EU’s laws, and the nature of the technologies themselves. Moreover, as pointed out below, for decades policy makers in Europe and beyond have been very much aware of and even legislated around generative AI (i.e. so -called computer generated works).

By approaching this issue from these various angles, we seek to furnish researchers and institutions more broadly with the confidence to push back on such assertions, thus making the most of these important flexibilities in EU law.

A Clear Legal Basis for AI Training

To begin with, the non-exhaustive definition of text and data mining in the CDSMD clearly envisages the use of a broad set of machine learning softwares:

Art 2 (2) ‘text and data mining’ means any automated analytical technique aimed at analysing text and data in digital form in order to generate information which includes but is not limited to patterns, trends and correlations.

Thus, as long as the computer model analyses copyright works (as an automated analytical technique) to generate an output the exceptions will apply. The same goes for the UK’s law which uses the broad term “data analysis”, already explored in a previous KR21 blog.1

The push-back from rightsholders perhaps is most acute when it comes to so-called “generative AI models.” The argument is very difficult to understand however given that in order to produce an output similar in style to the works one associates with the entertainment industry, as with any model that trains on data (whether for TDM, generative AI or any other use) the software must analyse the works in the first place, and therefore Arts 3 and 4 very clearly apply.

Moreover, it is this logic, one based on the technological process, which has led the European legislator to explicitly confirm in the AI Act that Arts 3 and 4 of the CDSMD apply when using general-purpose AI models, including generative-AI models.2

Recital 105

General-purpose AI models, in particular large generative AI models, capable of generating text, images, and other content, present unique innovation opportunities but also challenges to artists, authors, and other creators and the way their creative content is created, distributed, used and consumed. The development and training of such models require access to vast amounts of text, images, videos and other data. Text and data mining techniques may be used extensively in this context for the retrieval and analysis of such content, which may be protected by copyright and related rights. Any use of copyright protected content requires the authorisation of the rightsholder concerned unless relevant copyright exceptions and limitations apply. Directive (EU) 2019/790 introduced exceptions and limitations allowing reproductions and extractions of works or other subject matter, for the purpose of text and data mining, under certain conditions. Under these rules, rightsholders may choose to reserve their rights over their works or other subject matter to prevent text and data mining, unless this is done for the purposes of scientific research. Where the rights to opt out has been expressly reserved in an appropriate manner, providers of general-purpose AI models need to obtain an authorisation from rightsholders if they want to carry out text and data mining over such works.

Article 53 – Obligations for providers of general-purpose AI models

1.     Providers of general-purpose AI models shall: […]

 (c) put in place a policy to comply with Union law on copyright and related rights, and in particular to identify and comply with, including through state-of-the-art technologies, a reservation of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790;

From a legal standpoint therefore, a combination of the definition of TDM in the CDSMD, alongside its applicability to the deployment of general-purpose AI models as explicitly set forth in the AI Act, is enough to establish that the TDM exceptions apply when training and using general purpose models such as generative AI.  

If focussing on the outputs of a machine learning model, it also needs to be remembered that from a copyright perspective the only relevant issue is whether there is material similarity between a used input and an output. An output such as a robot moving its arm, a number, or a percentage of chance etc are not things regulated by copyright. Where the output generated involves something “creative”, the test the law applies is one of “substantial similarity”.

Analysis however shows that instances where the models have overfitted to the training data are rare.3 Furthermore, thinking about style transfer models, style is not something that is protectable – indeed, it is explicitly carved out from copyright in some jurisdictions by exceptions for pastiche.

Hardly a New Issue

Any claim that applicability of the TDM exceptions for AI could not be envisaged when negotiating the CDSMD is further undermined by the long history of copyright lawmakers reflecting on the generation and status of computer-generated works.

AI text and music generation, automated translation and other generative applications have been around for well over 35 years.

Although academic discussion on generative AI started even earlier,4 in copyright policy circles, legislative actions around computer-generated works goes back to the 1980s. The UK was the first country in the world to introduce a provision in copyright law for computer-generated works as far back as 1988, with Ireland following suit in 2000. In 1990, whether to accommodate such works into copyright law was actively being discussed at WIPO, and the following year, WIPO even held a worldwide symposium on IP and artificial intelligence at Stanford University including the showcasing of AI generated artwork from programmer-artist Harold Cohen.

Thus, even putting aside the explicit reference to general purpose models in the AI Act, given the long copyright history of discussing and even legislating for generative AI, the claims from some rightsholders that the TDM exceptions couldn’t have envisaged generative AI ring very hollow indeed.

The issue was also firmly on the table during the negotiations of CDSMD. In a January 2018 blog post (well over a year before it was adopted by the Parliament), Felix Reda – the Shadow Rapporteur in the European Parliament for the Green Party at the time – wrote regarding the proposed neighbouring right for press publishers:

“The lack of an originality requirement would mean that unoriginal longer texts published by news sites are covered, such as tables of football results or AI-generated texts, which are currently ineligible for copyright protection for good reason.” 

As one of the MEPs finalising the text of the Directive in trilogue, alongside Rapporteur Axel Voss, we can see that there was clear awareness of the application in the creative industry sector of generative AI.

Furthermore, the CDSMD itself formed part of the Digital Single Market (DSM) Strategy launched by the European Commission in 2015. As part of this initiative, the Commission’s 2017 mid-term review placed significant emphasis on the importance of AI, and therefore deep learning, training, generative AI and more. This connection was explicit in the European Parliament’s press release issued upon adoption of the Directive,  highlighting that these provisions were adopted

“… in order to contribute to the development of data analytics and artificial intelligence”.

Moreover,  the same year, we can see that when the European Council moved to introduce Art 4 into the Directive, it was couched in terms of “unleashing the potential of artificial intelligence, Big Data and innovation, which are important policy objectives of the European Union…”.

In Short…

Any claim that the TDM exceptions do not allow analysis of data in the context of deploying general purpose models or generative AI appears entirely erroneous. As highlighted above, there was a long-standing awareness amongst legislators of generative AI reaching back many decades. Even if this were not the case, in the EU, the explicit reference to the operativity of Art 4 of the CDSM Directive in the context of the AI Act should put the issue to bed once and for all.


Read other recent publications about copyright, research and AI:


  1. For an examination of this claim regarding s29A of the UK Copyright Act, see this KR21 blog. ↩︎
  2. See EU  AI Act. Recital 105. Art 53.1 (c) ↩︎
  3. Nicholas Carlini et al ‘Quantifying memorization across neural language models’ arXiv:2202.07646, 1 (referencing studies that find between 0.00000015% and 1% of training data may be remembered by the model and therefore risk an output being substantially similar to a model input.  ↩︎
  4. Milde Jr, K. F. Can a Computer Be an” Author” or an” Inventor”?’(1969). J. Pat. Off. Soc’y, 51, 378-404. ↩︎

3 June 2025

Related Content