Generative AI and Copyright - Challenges for Rights Holders and Users


This article provides an overview of copyright issues currently being posed in connection with generative AI under German copyright law.

The rapid technical development of generative AI has meant that creative content of all kinds can now be created with very little effort and in a very short time. Generative AI can be used to create not just all kinds of texts, but even images, programming code and videos.

Though the opportunities that generative AI offers are extensive, the problems and challenges related to the copyright for products that were created by or using generative AI are still unresolved and far-reaching. Both copyright holders* and users of copyrighted content should be aware of these problems and challenges so they can recognise the potential liability risks that arise in relation to the use of generative AI or the content it generates.

Artwork created using AI – US court denies authorship

An as yet unresolved, but fundamental question is who can own the copyright to the content generated by generative AI.

A US court recently took a position on this issue. On 18 August 2023, the US District Court for the District of Columbia issued a ruling in the case of Thaler v. Perlmutter. The subject of the case was a two-dimensional work of art created by a generative AI called the Creativity Machine, which was developed by American Stephen Thaler. Thaler then applied to the US Copyright Office for the artwork to be entered into the US Copyright Register.

The US Copyright Office denied the application on the grounds that the artwork was not created by a human being. The US District Court for the District of Columbia upheld this decision, emphasising that where no human influence was involved in the making of a creation by artificial intelligence, authorship must be denied. It also held that the fact that Thaler had developed the "Creativity Machine" himself did not entitle him to authorship of the work.

However, the court did not categorically rule out human authorship of AI-generated works. It clarified that authorship depended on the extent to which the creation of the AI-generated work was ultimately predetermined or influenced by a human being.

Possible conditions for authorship of AI-generated works to be approved

The question of when human authorship of an AI-generated work can be assumed is new legal territory. However, just as the US court in Thaler v. Perlmutter did not rule out human authorship of AI-generated works per se, authorship of such works in German copyright law also depends on the circumstances of the individual case, in particular on how the work created by the AI actually came into being and to what extent a human influenced the creation of the work.

Generative AI differs from traditional tools for creating copyrighted content in that the content is created by the AI largely "on its own" and at the push of a button. This is problematic not just under US copyright law, but under German copyright law as well, in that a work is only protected according to section 2 (2) German Copyright Act (UrhG) if it constitutes a person's intellectual creation. Products of chance or products found in nature generally do not enjoy copyright protection. Copyright protection is always conditional on a sufficiently controlling influence of a human being in the creation of the work.

The decisive factor here is likely to be the extent to which the creative work is already predefined in the prompt given to the generative AI and the extent to which this restricts its creative freedom. The more the result is predetermined by the prompt and the less creative freedom the AI had in generating the content, the more likely it will be that the prompt creator will be assumed to be the legal owner of the work. However, this always needs to be checked on a case-by-case basis and the product of the AI cannot always be clearly predetermined with a prompt. Especially in the case of image-generating AI, it is evident that entering the same (detailed) prompt multiple times often results in completely different images.

Are AI-generated works protected by rights related to copyright?

In addition to works, the German Copyright Act also protects certain economic, organisational and technical contributions ("related rights"), which do not necessarily have to be personal intellectual creations.

For various reasons, however, related (intellectual property) rights are only suitable for the products of generative AI to a very limited extent. For example, the related rights of a photographer under section 72 German Copyright Act (UrhG) requires that the work is a photograph or a product manufactured in a similar manner to photographs, which is not the case with computer-generated images.

Likewise, the related rights of a producer of a database under section 87a German Copyright Act (UrhG) does not protect the mere creation of data, meaning that merely creating such works does not already lead to the right of a producer of a database to these works.

The new related rights for publishers of press publications under section 87f German Copyright Act (UrhG) also only protects press publications that consist mainly of written works of a journalistic nature. However, journalistic texts generated by means of AI do not constitute works under section 2 (2) German Copyright Act (UrhG).

Copyright risk: Using copyrighted content to train generative AI

The developers of generative AI are faced with the question of whether they can use copyrighted content to train generative AI. It is clear that generative AI requires the use of large data sets, which very often include material protected by copyright. Developers often use the internet itself, with its enormous wealth of copyrighted content, to compile training datasets. However, this has now led to several court cases being brought against various developers of generative AI in the US.

Under German copyright law, there is a question as to whether the compilation and use of training datasets for the purpose of training generative AI is justified under the text and data mining limitations of section 44b and section 60d German Copyright Act (UrhG). The limitations allow protected works to be reproduced for the purposes of text and data mining, i.e. the compilation of corpora of different protected works for automated analysis in order to obtain information about patterns, trends and correlations.

However, applying these two limitations in the context of training data presents a number of problems. Whereas section 44b German Copyright Act (UrhG) can be invoked by anyone, section 60d German Copyright Act (UrhG) is only applicable in a non-commercial, scientific context. In addition, both limitations require lawful access to the protected content. It should also be noted that within the framework of section 44b German Copyright Act (UrhG) works cannot be reproduced if the holder of the rights has reserved the rights to the works under section 44b (3) German Copyright Act (UrhG), which in fact means it is necessary to obtain prior approval from the owner before using any legally protected content for training purposes. Finally, both limitations also provide for comprehensive requirements to delete data, which prevents protected works from being used extensively in the context of training data.

Against this backdrop, the only legally secure method of using copyrighted content as training data sets currently available is to obtain the appropriate licences and permissions from the holders of the rights.

Risk of infringement through the use of AI-generated content

Using content generated by generative AI may also infringe third-party rights.

Since copyright does not protect styles, everyone is free as a matter of principle to create works in the style of a particular artist. Generative AI can also be used in this way to create works that appear to be by a particular well-known artist. Care must always be taken to make clear that the work is not actually by the respective artist. If the impression is created and maintained that the work is by a particular artist, this may constitute an infringement of the artist's personal rights.

Furthermore, using content produced by generative AI always poses the risk that the content itself is an adaptation or reproduction of the content with which the generative AI was trained. Therefore, there is no way to rule out that in certain cases, such as when an image is rare and has been used as training data, this image is copied (unchanged) in the output.

Provisions in the draft AI Act on copyright

In the current version of the draft AI Act there are only a few provisions concerning copyright.

The draft stipulates in Article 28b (4) c) that operators of generative AI are required to document and disclose all training data that are copyrighted and which have been used in the training of the AI.

The draft regulation also clarifies that it is without prejudice to the directives on copyright previously adopted by the Union legislator. The questions about the protectability of AI works as well as the permissibility of using copyrighted content as AI training data will therefore continue to be addressed according to established and EU-harmonised copyright law even after the latest version of the AI Act enters into force (see our blog for more information on the regulation of generative AI through the AI Act).

The emergence of generative AI systems poses multiple copyright challenges to holders of rights and users of works

It remains to be seen how the courts in Germany will decide questions of authorship of AI works and the legality of using copyrighted content as training data. At present it cannot be ruled out that further regulations concerning copyright will be included in the AI Act. Holders of rights and users alike should both be aware of the risks associated with the use of generative AI systems and should always carefully check whether the respective AI work is protected by rights or whether third-party rights could be infringed.