The technical advances of the 21st century are truly amazing. However, ever new possibilities also entail the need to constantly develop new regulatory systems or adapt existing systems to the state of technological development. However, law develops far more slowly than technology, so that regulatory voids sometimes arise.
What are AI image generators?
In short, the corresponding programs can convert any text into an image. The user describes the desired graphic – no matter how absurd – and the generator develops an image based on the text. If desired, the results can depict a lion on a bicycle, Donald Trump on the seabed or even “simply” a lawyer at a desk. Complex descriptions of objects, patterns, colors, texture, style, etc. can also be specified (possibly in combination) and are applied by the AI. For this purpose, the programs are fed with data in the course of machine learning and trained to recognize and reproduce specific aspects. The concept has been improved since the 2010s, particularly with regard to the quality of the generated “artwork”. In the beginning, it was only possible to create blurry black and white images measuring around 2 centimetres. Nowadays, however, the AI products look high quality, have a wide range of colors and the image size can also be varied.
The question of the author
To date, there are no specific standards in the German legal system for the use of AI image generators. The question of whether the result of an algorithm is worthy of copyright protection has been discussed since the 1970s; however, AI generators take this discussion to a whole new level.
- Who is the author of the AI products?
The main problem is that AI creates “works of art without artists”, so to speak. The author is initially the person who has created a work. However, an author under German law within the meaning of Section 7 UrhG must always be a natural person – meaning that the AI itself is not an author. The operator or developer of the program or the respective user who provides the input for the work to be created can also be considered.
In principle, the AI has fewer problems with creating the graphics and greater difficulties with actually understanding the text. . It would be conceivable to assume authorship for the user “if the influence of a human is so decisive that the result of the design process can still be attributed to them.”
The user can steer the result in the desired direction by adding various criteria or by reformulating it. However, it is difficult to determine or measure the actual creative output of the user as opposed to that of the AI.
- Can input be protected by copyright?
It can also be considered whether the input command falls within the scope of protection of the Copyright Act. However, it seems questionable what requirements must be placed on the character string entered. If, for example If, for example, not every short message on Twitter has a sufficient “creative level”, this cannot generally be assumed for every word combination in the input field of an image generator.
Without comprehensive standardization, the question of the author remains a case-by-case consideration.
Can AI image generators infringe existing copyrights?
Unlike authorship, no human contribution is required for an infringement. The AI uses so-called datasets to generate images that are as realistic as possible. These are a collection of images that the AI uses to learn what certain objects should look like. AI can therefore only deliver realistic results on the basis of these datasets – for example, if you enter the command “Apple in the style of Banksy”, the AI must first have received training data from which it could learn how a) what an apple looks like and b) the style of Banksy’s street art. In principle, no copyright can be claimed for a specific style. However, if you give the AI the command “draw a cartoon mouse in red shorts”, for example, it will almost certainly generate an image that is very similar to that of the Disney character Mickey Mouse. The question now is whether there is a copyright infringement in the fact that images of Mickey were used as training data for the AI.
So far, however, it has been difficult to draw concrete and conclusive conclusions about the components used from the generated image – copyright infringement is therefore not easy to prove.
It is particularly problematic that licensing the data used does not appear to be practicable, as the amount of data used to train the AI is so huge that paying license fees would drive providers to financial ruin.
Other intellectual property rights may also be affected. If the AI learns to associate DHL trucks with the term “Postauto” and then generates an image with the word DHL on it, this could, for example, infringe the German Trademark Act. the MarkenG could be infringed. The faces of other people could also be included and infringe their personal rights. This would also have to be assessed critically in terms of data protection law,
The legal situation is currently still so uncertain that even the photo agency Getty Images has imposed a ban on the use of AI-generated images on its own portal due to legal concerns.
But what can authors do to ensure that their images are not used for AI training purposes? In principle, it is possible to include a so-called reservation of use, e.g. in the terms of use and/or on your own website. However, it is still uncertain whether this is sufficient and what exactly such a reservation should look like. Here too, the legal situation is currently still extremely unclear, and more detailed explanations on this are to be expected from the courts in the future.
However, the use of third-party copyrighted works for scientific purposes – i.e. theoretically anything that does not pursue commercial interests, especially in the field of AI – is expressly permitted. The author cannot defend themselves against this, at least not at present.
The haveibeentrained.com website can be used to track whether a work has been used for AI training.
No legal certainty for software operators
In a separate article, you will find an overview of the currently most relevant image generators, how they obtain their training data and how the transfer of rights to the generated result is structured. Most of the information on this can be found in the terms of use of the programs, although these are always quite extensive and not always obvious to find, but do not guarantee any real legal certainty with regard to any copyright infringements. Instead, the risk of any infringements is regularly passed on to the user.
The overall picture shows that most software operators themselves do not know for sure what the legal situation is and how the terms of use must be structured. The regulations are superficial and statements regarding authorship are only made vaguely.
Read the AGB
Before you use one of the programs provided for this purpose or an image created with the help of such a program, you should always read the terms and conditions carefully and, if necessary, even ask the operator who owns the rights to the works in question.
Ultimately, the specific legal position of the program user depends on which legal system is applied. For example In the USA, for example, a complete transfer of copyrights is possible, whereas the German legal system does not provide for this.
It is far from clear where developments in this area will ultimately lead. The commercial benefits are huge and progress in program development is rapid. Without a relevant legal definition by the legislator, many (legal) questions remain unresolved.
Thedocumentary “The AI that creates any picture you want, explained” offers a worthwhile insight into how AI image generators roughly work.
Focus on AI & copyright
In our thematic focus “AI & Copyright“, we deal with the copyright aspects of current developments in the field of artificial intelligence. So far, this series has included introductions to the applications ChatGPT and the use of AI art, an overview of the terms of use and training data of popular AI image generators and the Mandelbrot debate in the context of the legal discussion. We shed light on the concerns and fears of press publishers in the article“Ancillary copyright and AI” and explore the fundamental question of whether AI can be an author in a separate article.