Voice assistants, smart homes, autonomous driving, AI is omnipresent and, at the latest since ChatGPT, more present than ever before. Exciting and fascinating as AI may be, it also brings legal challenges. Questions arise, in particular, in the area of intellectual property, or more precisely, in the area of copyright law. How do traditional copyright law, designed for analogue works, and digital output generated by an AI fit together? The following article will provide an overview of the challenges for copyright law in connection with AI-generated output.
Copyright protection and AI output
According to German copyright law only a “work” within the meaning of Section 2 (2) of the German Act on Copyright and Related Rights (Urheberrechtsgesetz, UrhG) is eligible for copyright protection. Consequently, AI output must be categorized as a “work” to enjoy copyright protection. Section 2 (2) UrhG defines a “work” as an author’s own intellectual creation, whereby such an author’s own intellectual creation requires human creative activity.
Within this legal framework, two main scenarios arise concerning AI-generated output:
First, the output generated purely by AI must be considered. That is the output in which all essential design decisions are made by AI. Generally, this output cannot be protected by copyright. The reason for this is that the output was not created by a human being, but by an AI. The requirement of an “author’s own intellectual creation” is not met as a result.
In contrast, the case gets more complicated if both – a human being and an AI – are involved in the creation of the output. The decisive factor here is if the human contribution to the final design of the output is sufficient to regard the human being as the creator. What is decisive in this context is whether the AI is used merely as a tool to implement design decisions developed by the human creator – like a paintbrush or a chisel – or whether the control over the design process and outcome remains with the AI. If the latter is the case, an “author’s own intellectual creation” does not exist, with the consequence that the resulting output is not protected by German copyright law.
This legal situation leads to several challenges for copyright law:
Number one is the question of when an output can be attributed to a human being and when to an AI. Consequently, it must be decided what criteria and standards will be used to assess the amount of the human participation in the output generated by AI. The second challenge is a mere practical one. How can the amount of the human contribution to the output be determined without any knowledge of the process of its creation? As only human creations enjoy copyright protection, there might be a strong incentive not to disclose the AI-generated share.
Finally, it must be clarified how to deal with output that is not a “work” and thus does not establish any copyright. Ancillary copyrights such as the database producer’s right (Section 87a UrhG) could come into play here. However, to develop a separate protection system for such output de lege ferenda might be the preferable way.
Text and data mining
For an AI to “come to life”, it is necessary for its underlying algorithm to be able to train its performance with the help of large data sets.
For this reason, it has to be assessed whether the use of data sets for training AI applications is legally permissible. In many cases, training requires the use of information contained in photos, texts, videos and other data sets, which may themselves be subject to copyright protection. This means training AI applications may interfere with other people’s copyrights. For example, corresponding data must be uploaded into the computer’s main memory for each automated analysis. This results in duplications and possibly also adaptations, which may infringe the author’s right of reproduction (Section 16 UrhG) or adaptation and transformation (Section 23 UrhG).
Until June 2021, permission from the author was required in Germany for the use of such data, unless the reproductions were made for the purposes of scientific research (Section 60d UrhG, former version). Consequently, permission from the author had to be obtained for commercial text and data mining. Given the amount of data required, this was effectively impossible.
To remedy this problem, promote technical innovations and at the same time create a secure legal framework for AI applications for companies, the European legislator introduced a provision for machine learning in 2019 with the DSM Directive (RL (EU) 2019/790 of 17 April 2019). The German legislator implemented this in Section 44b UrhG, which came into force on 7 June 2021.
According to Section 44b UrhG, “it is permitted to reproduce lawfully accessible works in order to carry out text and data mining. Uses […] are permitted only if they have not been reserved by the rightholder”. Thus, there is (general) legal permission to collect copyrighted works and use them to create training data. Therefore, regulations are no longer limited to text and data mining for scientific purposes, but now also allow text and data mining for commercial purposes (BT-Drs. 19/27426, 87).
However, three essential requirements must be observed for permissible commercial text and data mining:
- Only lawfully accessible works may be used for text and data mining. This is the case if the work can be accessed without the user committing a copyright infringement.
- The training data must be deleted when it is no longer needed for text and data mining. As a result, storage of data beyond a specific AI project is excluded, for example.
- The restriction of Section 44b UrhG does not apply if the rightholder has reserved the right of use. Consequently, the exploiter does not have to ask for permission, but the rightholder must act if they want to prevent any potential use for text and data mining.
Infringement through AI-generated output
Users of AI applications should ask themselves whether the output generated by AI could constitute a copyright infringement.
A copyright infringement can in principle also be committed by an output that is not considered as a “work” according to German copyright law and thus does not qualify for copyright protection itself. Whether the AI-generated output infringes any third party’s copyrights essentially depends on whether the output constitutes a reproduction (Section 16 UrhG) or an adaptation (Section 23 UrhG), or whether the use is not copyright-relevant at all.
The demarcation between a prohibited adaptation and a permitted copyright-irrelevant use had already arisen in many ways in the “analogue world”. Nevertheless, this demarcation is still difficult. The decisive factor, which also applies to AI-generated output, is the content-related distance between the output and the original work. This must be assessed on a case-by-case basis. If the output generated by AI shows a sufficient content-related distance to the original work, the consent of its author is not required. Otherwise, it is a case of an unauthorized adaptation, which constitutes a copyright infringement unless the author has explicitly given their consent.
If it has been established that an unauthorized reproduction or adaptation is involved, the following applies to the relationship between the user and the author of the original work. The author is entitled to injunctive relief against the user under Section 97 (1) UrhG. In this context, the user has committed a copyright infringement if they make the output generated by the AI publicly accessible. Furthermore, a claim for damages under Section 97 (2) UrhG could be asserted against the user – if the user’s behavior has been negligent or intentional.
Conversely, the user cannot be held responsible for errors of the AI that are unrecognizable to them. Nevertheless, users can rarely claim that they were unable to assess or be aware of the functions of the AI but use the output anyway. Therefore, it is the responsibility of the user to review the content generated by AI. If they do not do so, they will usually have committed a breach of due diligence, and the author may hold them responsible for any infringements caused by using an AI.
Summary
AI-generated output imposes a number of challenges on copyright law. However, not all these challenges are new. The questions of demarcation between a prohibited adaptation of a copyright protected “work” and a permitted copyright-irrelevant use, or the use of computer programs as technical support, are nothing new to copyright law. Still, AI-generated output will also have to be evaluated here on a case-by-case basis. This requires appropriate standards to be developed by case law. Other challenges seem to have been solved, at least for the time being – for example, the permissibility of text and data mining for commercial purposes (even without the permission of the authors concerned). But there are also some new challenges to be addressed. In particular, these concern the question of how to deal with AI-generated output that is not copyrightable, as well as the question of how to determine how large the respective human or technical contribution to an AI-generated output is.