Mass claims and collective redress as a result of the training of Artificial Intelligence based on user generated content?

To train algorithms used by artificial intelligence, large batches of data are fed to the algorithm. This data often includes user generated content (such as pictures, texts and videos) protected by copyright. The use of this data may constitute copyright infringement. For the individual user it may be relevant to investigate whether users could bundle forces and start a mass claim against these uses.

Representative actions are already relatively common in Europe. For example, in Germany, the GEMA – the German collecting society for musical performing and mechanical reproduction rights – sued OpenAI and claimed that ChatGPT was trained on lyrics of nine well-known authors without permission.[1] Lastly, in the UK, Getty Images – a global provider of licensed photographic content – sued Stability AI and alleged that Stability AI used its images without consent for AI training.[2]

It is only a matter of time before we can expect a similar case in the Netherlands, where, in addition to representative actions, another legal framework may be relevant in this regard. In January 2020, the Netherlands introduced a ground-breaking piece of legislation known as the Wet afwikkeling massaschade in collectieve actie (“WAMCA”), which translates to the Settlement of mass damages in collective action Act. The WAMCA provides a more efficient and effective framework for handling mass damage claims through collective action.

In this article, we will investigate whether the WAMCA offers an effective tool for collective redress as a result of the use of copyright protected works for the training of algorithms used for artificial intelligence.

Copyright ingringement as a result of the training of AI

Nowadays, artificial intelligence is used by almost anyone and everyone, and artificial intelligence is used for several purposes. For example, artificial intelligence can be used to detect harmful content,[3] to recommend and/or moderate content,[4] for smarter advertisement targeting based on user behavior and preferences,[5] and of course to deliver hyper-personalized content feeds.[6]

To develop an algorithm that serves all these purposes, the algorithm must be trained. For example, the spam filter in an email system is trained by feeding it large quantities of emails marked as either “spam” or “harmless”. The algorithm analyzes all emails and discovers, for example, that 80% of the emails that contain the word “free” are marked as spam. This pattern (and many others) is then used to determine whether an email is spam or harmless.[7] The algorithm is usually fed not only data on user behavior (such as likes, engagement, etc.),[8] but also user generated content.

a) Does the use of user generated content constitute an infringement?

Most of this user generated content is protected by copyright. The threshold for copyright protection in the Netherlands is relatively low. The work must be sufficiently original (meaning it cannot be derived from previous works of others) and it must reflect the author’s personality (meaning the content needs to be creative and include sufficient free and artistic choices). Applying this threshold to for example pictures, creative choices can be made with regard to the chosen background, the subject’s pose, and the lighting.[9] When taking a portrait photograph, the photographer can choose the framing, the angle of view and the atmosphere created.[10] Finally, when selecting the picture, the photographer may choose from a variety of developing techniques and make final touches.[11] If sufficient of these creative choices are made, the picture is protected by copyright. The same applies to for instance videos and texts.

The question is whether use of user generated content for training purposes constitutes a copyright infringement.

This depends on whether or not this use is a relevant reproduction. Although some discussion remains – for example because reproductions under copyright law imply the use of the work as a work, whereas in this case reproductions are made to train the model and to generate a different output – training a model by feeding it copyright protected works most likely constitutes a relevant act of reproduction.[12] This is also the approach applied by the EU.[13] Because the reproduction right is the exclusive right of the author, an exception or limitation is necessary to avoid copyright infringement.

The most relevant exception is the exception for commercial text and datamining. Pursuant to Article 2(2) of the DSM Directive, text and data mining is “any automated analytical technique aimed at analyzing text and data in digital form in order to generate information which includes but is not limited to patterns, trends and correlations.” Training algorithms to identify patterns and optimize the algorithm by feeding it (user) data is therefore a form of text and data mining. In the Netherlands, the exception for commercial text and data mining of Article 4 DSM was implemented in Article 15o of the Dutch Copyright Act (“DCA”). The following conditions apply:[14]

• Access to content must be based on open access policies or contractual arrangements (“lawful access”). Works that are freely accessible on the internet are also works to which lawful access exists and which may therefore be copied for the purposes of text and data mining for scientific research.[15]
• If the copyright has been expressly reserved by their right holders in an appropriate manner, this work may not be copied for the purpose of text and datamining without the explicit prior consent of the right holder.[16] This opt-out must be applied in a machine readable format, such as in the metadata where the conditions for use of a website are indicated.[17]
• Article 15(2) DCA: the reproduction may be retained for as long as necessary for text and data mining. Once text and data mining has been completed, the reproductions made for this purpose must be deleted.

If the exceptions for the commercial text and data mining exception are not met, the use of copyrighted material to train algorithms probably constitutes a copyright infringement. In addition, it may be argued that knowingly placing an AI model on the Dutch market which was trained on copyright protected works, leads to an unlawful act based on Article 6:162 of the Dutch Civil Code (“DCC”).

b) Can an infringement or unlawful act be enforced?

If an infringement or unlawful act as a result of the use of copyrighted materials to train algorithms is suspected, the next question is whether this infringement or unlawful act can be enforced. The individual author behind the user generated content may think it is not worth it to pursue this matter and initiate legal steps.

Because training an algorithm requires an enormous amount of data, it is possible that all of these batches of data constitute batches of infringements. The proprietors of the copyrighted data that is used, may decide to bundle forces and start a mass claim against the user. In the Netherlands, representative actions are already relatively common. For instance, BREIN often litigates against alleged infringers by lodging collective claims. BREIN for example successfully blocked The Pirate Bay, lodged a collective claim regarding the supply of illegal e-books, and regularly initiates proceedings against large uploaders. Question is whether collective redress using the WAMCA would also offer a solution.

The requirements for collective redress - WAMCA Framework

 a) Introduction

The WAMCA, which entered into force on 1 January 2020, enables a representative interest organization to seek collective compensation on behalf of a group of injured parties in cases of mass damages. The statutory provisions introduced by the WAMCA – particularly Article 3:305a DCC – are designed to facilitate a more efficient and uniform handling of mass damage cases, by allowing common issues of law and fact to be resolved in a single set of proceedings. In the first phase of the proceedings, the admissibility of the claim and the claim organization is assessed. After an oral hearing, the court renders a decision on the admissibility only. At this stage, the court also designates the exclusive representative (in case several claim organizations start the same collective action), clearly defines the group of affected persons and provides the members of that group with a period of at least one month to declare that they do not wish to be bound by the outcome of the final judgment (so-called opt-out).[18] Subsequently, the substantive phase of the proceedings commences, which culminates in a final judgment or a collective settlement.

The WAMCA provides opportunities for collective damage claims across various areas, such as consumer and investor matters, privacy and data breaches, competition law, and environmental and product liability. Although the WAMCA can be applied in a wide range of fields, its use could become particularly relevant in the context of the use of copyright-protected works for training algorithms. The key question in such cases is whether this type of infringement is suitable for aggregation and whether a collective action is appropriate to seek compensation for damages.

b) Statutory requirements

Under the WAMCA, a collective action may be initiated by a representative organization (a foundation or non-profit association) with a sound governance structure, relevant expertise, and adequate financial resources. The interests to be protected must be sufficiently homogeneous to allow for efficient handling in a single procedure (commonality requirement). In addition, there must be a close connection to the Netherlands, for example, because the majority of the injured parties reside in the Netherlands, the defendant is established there, or the damage occurred within the Netherlands.

The commonality requirement

The aforementioned commonality requirement means that the interests or claims of the group must be suitable for aggregation in a single procedure, without the court needing to examine in detail the individual circumstances of each injured party.

In 2010 and under the predecessor of the WAMCA, the Dutch Supreme Court ruled that the commonality requirement is satisfied if:[19]

“[…] the interests for the protection of which the legal claim is brought are suitable for aggregation, so that efficient and effective legal protection for the stakeholders can be promoted. This allows a single procedure to address the disputed points and claims raised by the legal action, without the need to consider the specific circumstances of each individual stakeholder.”

The focus is not only on whether the interests are suitable for aggregation, but also on whether the individual rights at issue in the procedure are appropriate for collective assessment. The claims brought must protect similar interests of injured parties. For these claims to be aggregated, they must be capable of being addressed in a single procedure without taking into account the specific circumstances of individual parties.[20] Although the focus should be on the interests of individual injured parties who could have brought a claim themselves, general interests may also fall, indirectly, within the scope of Article 3:305a DCC.[21] These interests must then be translated into a traceable benefit for individuals, such as the interest of all rights holders in the protection of copyrighted works.

In this case, the commonality requirement means that it is not necessary to determine for each individual work of a rights holder which specific portion was used, but it must firstly be established that each work is protected in accordance with the requirements and secondly, whether each work was used without permission as part of the training data, which could complicate the procedure.

Aggregation may also be difficult when specific damages for each rights holder need to be determined, as this requires individual inquiry and may result in the inadmissibility of the collective claim. In practice, this can, for example, be addressed by creating different categories or classes of claimants, within which the claims and circumstances are sufficiently homogeneous to allow for joint consideration by the court.

The sufficient link-requirement

Furthermore, the claim that forms the subject of the collective action must have a sufficiently close connection to the Dutch legal sphere, in order to prevent that Dutch courts are flooded with collective action that lack a real relationship with the Netherlands.[22] The requirement is satisfied if:

1. the interest organization demonstrates that the majority of the injured parties whose interests are sought to be protected by the legal claim have their habitual residence in the Netherlands, or
2. the defendant is domiciled in the Netherlands and additional circumstances indicate a sufficient connection to the Dutch legal sphere, or
3. the event(s) to which the legal claim relates occurred in the Netherlands.

This requirement is interpreted strictly in case law. For example, in the case TPC v. Oracle and Salesforce, the District Court of Amsterdam held that a close connection did not exist, because the data leak occurred in the United States and no more than eighteen injured parties resided in the Netherlands.[23] In the case concerning sand extraction in South Sulawesi, carried out by an Indonesian subsidiary, the court likewise concluded that a close connection was lacking.[24] The fact that Boskalis was established in the Netherlands, had engaged with the interest organization, and that the Dutch embassy was involved, was insufficient to allow a collective claim on behalf of the victims of the sand extraction.

The cross-border and digital nature of training algorithms raises specific challenges in this regard. AI models are often trained on international datasets containing works from multiple jurisdictions, making it difficult to determine what proportion originates from Dutch rights holders. Many AI companies operate globally, through subsidiaries, branches, or solely digital presence. A defendant may be headquartered in the Netherlands but may still use works of Dutch rights holders without their permissions. Furthermore, it remains unclear how many and what types of connections are required, or when harm is felt globally but also partially in the Netherlands.

Conclusion

Cases such as those brought in Germany and the UK illustrate that (collective) litigation over AI training practices and possible copyright infringement is no longer hypothetical. Given the rapid development of AI systems and the growing scrutiny regarding training data, it is only a matter of time before a similar case arises in the Netherlands. When it does, the WAMCA can provide an interest organization with a powerful tool to pursue collective claims for damages. However, in the context of IP and AI cases, it remains an open question whether both the commonality requirement and the sufficient connection to the Netherlands are adequately satisfied.