AI-based applications (e.g. large language models such as ChatGPT) are trained with vast amounts of data, often including copyrighted materials. But to what extent is it actually permissible to use copyrighted content in AI training? This question is currently the subject of a controversial debate, which is bound to keep the courts busy in the coming years. The New York Times has been the first major player to file a lawsuit on this issue, seeking to block the use of newspaper content for training the large language model ChatGPT.

In Austria and the EU, there is a statutory provision (§ 42h of the Austrian Copyright Act) that permits the use of copyrighted content for the purpose of text and data mining (TDM). There are good reasons to believe that the training of a generative AI also qualifies as TDM and therefore does not require the authors’ consent. If authors want to make sure that their content is not used to train a commercial AI applications, they can prevent this by issuing a (machine-readable) reservation to opt out of AI training. Whether this legal situation is suitable for promoting the development of AI applications and balancing the interests of AI operators with those of content authors needs to be investigated in more detail.

Copyright protection of AI-generated content

The question also arises as to whether texts or images generated by AI applications themselves are subject to copyright protection. Whether or not AI users can come into conflict with copyright law when using a text created using ChatGPT ultimately depends on the answer to this question. Even though this issue is not explicitly regulated in the Austrian Copyright Act, an overall assessment of the legal situation shows that currently only human creations are covered by copyright protection.

The key question here is whether specific texts or images generated with the help of AI applications can be qualified as human creations. This is the case if at least one human being was able to foresee the elements eligible for copyright protection to a sufficient extent (in the case of texts, this refers in particular to their linguistic structure, and in the case of images, to their visual characteristics). In general, this is unlikely to apply to the developers of AI applications, due to the complexity of these systems.

When it comes to the users of AI applications, on the other hand, an important distinction must be made: Simply entering a question in ChatGPT is not sufficient to obtain a copyright to the output text. However, if users enter prompts that outline the output in a very specific way, it is not categorically out of the question that the AI output may be eligible for copyright protection. A recent court ruling issued in China has granted copyright to a person who used an AI-based image generator and kept adapting their prompts until the AI eventually generated an image that closely reflected their initial vision. This shows that in individual cases, it is conceivable that AI users may obtain a copyright on AI-generated material.