Copyright © Media Law International 2023. All Rights Reserved.                                                                                                              

Specialist Guide to the

Global Leaders in Media Law Practice

      M E D I A   L A W   

      I N T E R N A T I O N A L   ®


Regarding the question of possible copyright infringements in connection with the use of copyright protected works for AI training, the first and probably most difficult question is which copyright-relevant uses of such works are taking place during AI training. This depends above all on the exact sequence of the technical processes.

First of all, it must be assumed that for the analysis of works protected by copyright, be they images, music or software programs, (at least temporary) reproductions must take place in any case for the analysis of the work and the associated metadata. A reproduction of a work however would only be possible with the permission of the author (Section 16 German Copyright Act) or if copyright law is providing an exception.

Transient copies

Insofar as copies of works are indeed purely temporary acts of reproduction without independent economic significance, the reproduction could already be permissible under Section 44a German Copyright Act, which is the implementation of Article 5 (1) of Directive 2001/29/EG (‘Info-Soc Directive’). However, it seems difficult to argue AI applications would not have an independent economic significance, as they are usually used in a system that is making available content to users (at least also) against payment, it will probably not be possible to invoke this restriction.

Text and data mining

However, if the reproduction was made solely for the purpose of so-called text and data mining, it could fall under the legal restriction of Section 44b German Copyright Act (text and data mining), the implementation of Article 4 Directive EU 2019/790 (‘DSM Directive’) or under the legal restriction of Section 60d German Copyright Act (text and data mining for the purpose of scientific research), the implementation of Article 3 DSM Directive.

According to Section 44b (1) German Copyright Act, "Text and data mining" is "the automated analysis of single or several digital or digitized works for the purpose of generating information, in particular patterns, trends and correlations." If the purpose of the reproduction is limited to this type of analysis to obtain information, then the reproduction can take place without the permission of the rightsholder. If the text and data mining is initiated for purposes of scientific research, as defined in Section 60d German Copyright Act, then no further requirements need to be met and the reproductions may also be made publicly available for joint research or quality testing within limited research groups.

The situation is different for text and data mining that pursues commercial purposes. Here, reproductions are only permitted if the reproduction was made from a lawfully accessible work for text and data mining (Section 44b (2) German Copyright Act) and also only if the rightsholder has not explicitly reserved his rights to text and data mining reproductions (Section 44b (3) German Copyright Act). According to the wording in the German Copyright Act, a reservation of use in the case of works which are available online is effective only if it has been made in a machine-readable format. The purpose of the machine-readable format is to be able to reliably indicate an opt-out to automated crawlers and, therefore, to respond to market realities.

As pointed out above, the text and data mining exception of section 44b was newly inserted into the German Copyright Act by the Act on the Adaptation of Copyright Law to the Requirements of the DSM Directive dated 31 March 2021, which means that operators of commercial AI systems can create reproductions for a text and data mining without the permission of rightsholders only since the new law came into force on 7 June 2021. Therefore, when assessing a particular situation, the question always arises as to the period of time during which reproductions for text and data mining have taken place and whether the rightsholders concerned have already declared an effective opt-out with regard to text and data mining during this period.

Three-step test

A further restriction in the application of exceptions and limitations may be imposed by the so-called three-step test, Article 5 (5) of the InfoSoc Directive. According to this European law requirement, "exceptions and limitations may be applied only in certain special cases where the normal exploitation of the work [...] is not impaired and the legitimate interests of the rightsholder are not unduly prejudiced." Accordingly, to the extent that the reproductions made pursuant to Section 44b of the Copyright Act result in the normal exploitation of the work in question being impaired and the interests of the rightsholder being unduly infringed, the exception might not apply.

AI generated images can be adaptations

The technical process implemented for an AI training and the application of the trained AI must be further reviewed to determine whether further steps might require the consent of the rightsholder. Specific uses could concern the right to adaptations.

According to Section 23 (1) sentence 1 German Copyright Act, adaptations or other transformations of a work may only be published or exploited with the consent of the author. This means that the mere production of an adaptation is basically free, only the publication or exploitation of an adaptation is subject to consent. For example, making an AI-adaptation of a work available to the public would require consent of the rightsholder. However, an adaptation requiring consent does not exist if the newly created work maintains a ‘sufficient distance’ from the used work, Section 23 (1) sentence 2 German Copyright Act. Such distance can be assumed to be sufficient if the individual character derived from an existing work 'dilute' to such an extent in comparison to the individual character of the new work that the existing work is no longer recognizable or only rudimentarily recognizable. With AI creations the distance might very often be sufficient, but probably not always. Especially if a user strongly refers to a certain motif of a certain artist in a prompt, images may be created that may fall under the concept of an adaptation.

Test cases and legislation

It remains to be seen whether the current copyright law is suitable for the new methods and technologies that have developed in connection with artificial intelligence. Courts will have to deal with the conflicting interests of rightsholders, technology providers and users. This area of tension is not new, however, and as with other new developments, such as the digitization of works, the making available of works via the Internet, the technically complex sharing of works in file-sharing networks, the hosting of works on sharehosters, or the use of works in supposedly private networks, the courts and, if necessary, the legislature will try to work out appropriate solutions here.

The EU AI Act announced for the second half of 2023 could be a first step in this direction.


Assessment of AI Training Under German Law

Written by Dr Ursula Feindor-Schmidt, LL.M.


We are taking a more in-depth look at the question of how the use of copyrighted works for AI training has to be evaluated from a European perspective, and here in particular from a German perspective. Due to their comparatively efficient and inexpensive proceedings combined with a high level of expertise, the German courts have always been a popular venue for test cases when it comes to questions of liability for copyright infringement. International rightsholders have used the advantages of German courts already several times in high level landmark cases, e.g. against the largest international sharehosting and streaming platforms (Rapidshare, Uploaded, YouTube).


  Middle East and North Africa

  North America


  Western Europe

  Central and Eastern Europe


Dr Ursula Feindor-Schmidt

Fachanwältin für Urheber- und Medienrecht (specialised lawyer for copyright law and media law)

Dr Ursula Feindor-Schmidt, LL.M. is senior partner at the specialised Media Law boutique Lausen Rechtsanwälte. She has a doctorate and a master’s degree in IP law and a specialised lawyers degree for copyright and media law. She is consulting national and international clients with regard to their digital strategies and is entrusted with high level test cases on platform liability and copyright compliance. She is also a specialist in the developing field of artificial intelligence, in deploying AI for legal tech and in AI regulation. She is a regular lecturer and expert speaker at national and international events, eg for Akademie der deutschen Medien, the International Publishers Association, Rise of AI or the Fordham IP Conference.



  Firm Rankings

  Enhance Your Coverage

  About MLI

  Order a Copy

  Contact Details