What you should know before using AI tools

by on 2. May 2023

Since the recent releases of free applications accessible to everyone that work with artificial intelligence (AI), not a day goes by without reports about this new technology and its possibilities. Experts and the media agree that this is the next big step in technology development, comparable to the general accessibility of the Internet or the introduction of the iPhone.

Currently, the focus of attention is on generative AI applications such as Dall-E 2, Midjourney, Stable Diffusion and Chat-GPT which are creating text, images and audio in seconds. And the results are sometimes almost indistinguishable from works created by a human.

AI strategy for publishers
For publishers and other media companies, these new technologies offer new opportunities but also risks. It is therefore advisable not to leave the use of AI to chance, but to proactively deal with various use cases and to lay down an internal strategy and maybe guidelines for use of AI by employees,” advises Dr. Ursula Feindor-Schmidt, a specialist in copyright and media law. In the following case studies, Dr. Ursula Feindor-Schmidt, LL.M. analyzes topics that are highly relevant to the use of AI, especially for publishers.

The use of images available on the Internet for an AI training
Copyrighted images created by publishers and their employees are available in masses on the Internet, e.g. in preview images of covers and book content on online sales platforms such as Amazon. By its own account, the “underlying dataset for Stable Diffusion is the English language subset 2b of LAION 5b https://laion.ai/blog/laion-5b/, a general crawl of the Internet created by the German non-profit LAION.” To the extent that copyrighted images are used for training AI applications, the question arises as to whether such use is lawful.

Whether the use of works that are accessible on the Internet but not specifically licensed is lawful is already the subject of several lawsuits in the USA and UK. For example, several artists, as well as the image agency Getty Images, have filed suit against the company operating the Stable Diffusion application, Stability AI, based in London. According to European law, the use of the publicly available image files could be permissible reproductions for a text and data mining. In the scientific field, such use is made possible by the exception of § 60d UrhG. Even in the commercial sector, reproduction for text and data mining is permissible under Section 44b UrhG if the works are lawfully accessible and if the copyright holder has not expressly reserved this right (possibly in machine-readable form). The question still to be clarified here, however, will be whether the use of the available images is exhausted in such a reproduction for text and data mining or whether other copyright-relevant uses are made here or whether the barrier can actually be used to enable business models without licensing the works required for this purpose, which substitute for the exploitation of the work. In any case, illustrators and publishers are advised to declare an (effective) reservation of rights if they want to prevent text and data mining of their works.

The use of AI-generated images on covers and in books
For publishers and their employees, on the other hand, the use of AI applications can also make work easier. Can publishers use AI-generated images as cover images or to illustrate texts?

The conditions under which an AI application and its work results can be used depend primarily on the terms of use of the respective application. Many of the applications also release the work results for commercial use, but specify a certain reference. In some cases, the terms of use require a grant of rights with regard to the input (the prompt) and sometimes also the work results for the inclusion in the AI database. Insofar as entire images are entered into an AI application in order to have them processed, this reproduction by the user of the AI itself may constitute a use under copyright law for which the consent of the author is required. The same applies to the publication of an AI-edited version of an image or even of an entire work, such as the illustration of a novel, insofar as the copyrighted storyline is simply adopted. It would be conceivable, for example, for an AI to create a graphic novel from a novel. The publication of such an adaptation could be prevented by the author of the novel. If, on the other hand, the AI application is allowed to generate an image independently (without including an entire work) based only on a prompt, the work result itself is probably not protected by copyright, because the subject of copyright protection can only be a personal intellectual creation according to section 2 (2) German Copyright Act (also “UrhG”). According to supreme court case law, this presupposes an act of creation by a human being. Accordingly, a cover created by an AI would not receive any protection. This means that the publisher would also not be protected from an imitation of the cover.

The creating an audio version of a book by AI-generated voices
AI applications could also be useful for creating an audio version of a book by AI-generated voices. For example, larger platforms have been offering a low-cost variant to the human narrator of audio books for some time. The book is automatically converted into audio by an electronically generated voice. What do publishers need to consider in this context?

The conversion to audio of a text work constitutes an adaptation within the meaning of Section 23 German Copyright Act. In order to publish or exploit such an adaptation, including the production of a sound recording or making it publicly available via the Internet, the consent of the author of the spoken work is required. Often, these rights are referred to as text-to-speech rights. On the other hand, by entering voice samples that are only several seconds long, AI voices can be created that are almost indistinguishable from the original. Insofar as publishers are considering using existing sound recordings of speakers here, the general right of personality under Article 5 (2) of the German Basic Law (also “GG”) as well as rights of the recording producer with regard to making the original sound recordings accessible for the production of the AI voice must be taken into consideration in accordance with Section 85 (1) sentence 1 case 1 and 3 of the German Copyright Act. In addition, the author of the contents of the sound recording, e.g. a speech protected by copyright (§ 97 in conjunction with § 16 German Copyright Act) could also assert rights. If necessary, however, the user of the original sound recording could also invoke the restriction regulation of Section 44b German Copyright Act both for the use of the recording and for the use of the content of the recording if its requirements (use for a text and data mining, lawful access and no declaration of a reservation) are met.

The creation of publishing texts by AI
Other use cases would be “feeding” an AI with appropriately curated training data to have it write the next volume of a particular series. Here, too, caution matters:

  • The input of the pre-existing works into an AI application constitutes a reproduction of the work requiring consent; and
  • The new text generated by the AI may constitute an adaptation of the original works, the publication and exploitation of which would in turn require consent.

The same applies to the creation of summaries of books, for example for a blurb or an advertisement for the book, or the creation of books from a large number of individual works, for example a guide book from a quantity of magazine articles from a certain subject area. In each individual case, it is necessary to check which copyrighted works are used to what extent and whether these uses require consent or are (as a exception) free to use.

This article has been published first in German in buchreport magazine 05/2023.