Is text and data mining synonymous with AI training?
Eleonora Rosati · Journal of Intellectual Property Law & Practice · 2024
In late September 2024, the District Court of Hamburg delivered what appears to be the first judgment in Europe1 on the construction and application of the national transpositions of the text and data mining (TDM) exceptions found in Articles 3 and 4 of the DSM Directive.2 The Hamburg court ruled that LAION could rely on the exception found in section 60d UrhG (TDM for scientific research purposes). By this provision, Germany had transposed Article 3 of the DSM Directive into its own law. A fundamental aspect of the decision that deserves greater attention is that the analysis of the court is incomplete. As such, it may not represent good guidance for either concerned stakeholders or other courts in Europe faced with questions of unlicensed TDM and subsequent training of Artificial Intelligence (AI) models. Specifically (and likely because of how the plaintiff photographer pleaded the case), the court failed to consider that the TDM exception for scientific research would not cover all of LAION’s activities as described in the judgment itself, notably the circumstance—following the completion of TDM activities—that LAION made the resulting dataset publicly available for anyone to use and for any purpose, including commercial AI training. Like the corresponding EU provisions, the German sections considered by the court are exceptions to specified restricted acts under copyright and other rights: Like Article 5(1) of the InfoSoc Directive,3 section 44a UrhG (temporary acts of reproduction), which was held inapplicable, only encompasses acts of reproduction; Like Article 4 of the DSM Directive, section 44b UrhG (TDM), whose applicability in the case at hand—while doubtful—was deemed not necessary to decide, only encompasses acts of reproduction and, insofar as the sui generis database right is concerned, extraction; Like Article 3 of the DSM Directive, section 60d UrhG (TDM for scientific research purposes) is limited to acts of extraction and reproduction. As also discussed in greater detail elsewhere,4 none of the exceptions above extend to the doing of subsequent restricted acts, including those performed by LAION when it made the dataset for image-text pairs (resulting from its TDM activities) publicly available for free on its website. Assuming that the German court was correct in holding that the German equivalent of Article 3 DSM Directive would apply to LAION’s own TDM activities, following their conclusion, LAION performed restricted acts that are not within the scope of any TDM exception (whether Articles 3 or 4 of the DSM Directive). By creating a dataset and making it available to the public, LAION undertook at least an act of communication/making available to the public. Under EU copyright, since the 2014 Svensson decision5 and subsequent case law, the provision of a link to protected content can be actionable under Article 3 of the InfoSoc Directive and be so even if the link provider does not pursue a for-profit intention. Think of GS Media6 and the relevance of the link provider’s own knowledge, even where the link provider might have provided the link in question for non-profit purposes. More recently, in VG Bild-Kunst,7 the CJEU further confirmed that linking to protected content may be restricted not only through technical means (e.g., a paywall) but also—at certain conditions—through contractual terms. The judgment is problematic for other reasons too. Conflating the TDM exception with AI training reveals another problem in the reasoning, While the court correctly considered the need to construe the TDM exception in light of the three-step test (as per Article 7(2) of the DSM Directive), it erred when it used it to interpret the TDM exception in such a way as to allow the blanket training of AI models, opining that otherwise the TDM exception in Article 3 would become devoid of meaning. The three-step test in Article 7(2) is relevant to construe inter alia Articles 3 and 4 therein, not to include within their scope the doing of additional acts restricted by copyright that a research organization or others can perform. While there is little doubt, as also noted by the court, that when the TDM exceptions were adopted in 2019, AI training was already understood to be a possible application of TDM, it was an undue simplification on the side of the court not to consider that restricted acts other than those covered by the TDM provisions would be performed in the scenario at hand. An additional problematic aspect of the ruling is that, despite the remarks on rights reservation under Article 4(3) of the DSM Directive, the court seemed to consider the concept of ‘lawful access’ (which is a requirement inter alia under the TDM exceptions) as synonymous with ‘public accessibility’. These are different things: for example, posting an image on a website for everyone to see makes that image publicly accessible, but not necessarily also lawfully accessible, given that the person who posted it might have done it without the consent of the concerned rightholder. Cases like GS Media (linking to publicly accessible yet unlawful Playboy pictures) or Renckhoff8 (use of a photograph posted on a website only for use there) come to mind. The circumstance that the plaintiff’s watermarked preview photograph was publicly available on a website does not mean that it could be used by anyone for any purpose. Opining otherwise would also be contrary to the characterization of copyright’s exclusive rights as being preventive in nature and the prohibition of exhaustion of the right of communication/making available to the public (Article 3(3) of the InfoSoc Directive). In light of the foregoing, it is clear that one thing is the undertaking of restricted acts covered by applicable TDM exceptions; another is the doing of further restricted acts, including those that are propaedeutic to the subsequent training and offering of AI models. Hence, statements of the court like ‘whether the dataset … is also used by commercial companies for the training or further development of their AI systems is irrelevant because the research of commercial companies is still research’ are problematic in that they unduly simplify and overall misunderstand both (i) the steps needed to transition from LAION’s own TDM activities to the subsequent training and offering of AI models and (ii) their relevance and treatment under copyright. While the AI Act,9 as the court noted, recognizes the relevance of TDM to AI training, it does not say—as the court appeared to imply instead—that TDM is synonymous with AI training or that everything in-between TDM and AI training is covered by Articles 3 or 4 of the DSM Directive. To conclude: while the Hamburg decision offers some valuable guidance regarding TDM and AI training, including observations—like those concerning rights reservation—that might prove helpful in future cases, the judgment appears incomplete in some key respects and for failing to address—let alone answer—the question of the relationship between TDM and AI training. Above all, it is flawed in the part in which it does not acknowledge the limitations in the scope of Article 3 of the DSM Directive (as well as Article 4 therein). Further guidance is therefore needed in order to tackle the interplay between TDM and AI training correctly.