Machine-Generated Images and Copyright

(Note added in May 2024: I wrote this article in back in September 2022, i.e., before Gen-AI gained the traction it has today. Yet its findings are, as of now, still valid.)

These days, artificial intelligence is capable of creating stunning artwork. Run an image search for Stable Diffusion, Google Deep Dream, DALL-E, or BigSleep, and you may be amazed by what these tools can do.

But what about copyright?

Yes, I know. No one wants to read about that. But still, if you ever want to use such images for commercial purposes, you should pay attention to this aspect.

Stable Diffusion

Let’s look at Stable Diffusion as an example. Its release was announced by stability.ai in August 2022. A few months later, everyone, even those living cozy lives under a rock, has probably heard of it.

If you haven’t seen Stable Diffusion in action, you’re missing out. Head over to huggingface.co and enter a short, descriptive text of something you want to see. (Be patient: depending on system load, it may take some minutes to get the result.)

I will not go into the technical details here since they have been discussed all over the net, see e.g. here. Just briefly, Stable Diffusion works on a massive chunk of data. And this chunk of data that I want to talk about—namely the copyright issues lurking within it.

The data used by Stable Diffusion has been generated by training the machine on millions of images and pieces of text, the so-called LAION-5B dataset. Some of these images are still subject to copyright. This has given rise to concern (see e.g. Luke Plunkett, "AI Creating 'Art' Is an Ethical and Copyright Nightmare" at kotaku.com). Hence, my first question: Does the training of an AI such as Stable Diffusion infringe on the right-holders’ copyright?

Another issue—probably more relevant in practice— is the question of derivative work. For example, according to Art. 11 of the Swiss Copyright Law, the copyright holder has the exclusive right to decide if their work may be altered and if derivative work may be generated therefrom. Stable Diffusion and similar tools make it easy to do precisely that, i.e., they can generate images that are likely considered to be derivative work of copyrighted original work.

Let’s look at these two points in more detail.

Copyright and Training a Machine

As I said, Stable Diffusion and similar AIs are trained by feeding them a large set of images with related text. The images are typically copied from the internet to a local computer, scaled, cropped, and normalized before being fed to the algorithm that trains the machine. Such acts include the copying and modification of the original artwork.

Does this training infringe the copyright of the right-holders of the images and/or text?

In Switzerland, Art. 24a of the Copyright Law allows creating temporary copies of copyrighted work, but only if the "sole purpose is to enable a transmission of the work in a network between third parties by an intermediary or a lawful use of the work.” This is unlikely to apply when training an AI since such training goes well beyond the simple transmission of data.

Then, there’s Art. 24d of the Swiss Copyright Law, which allows the reproduction of a work for scientific research. The Comments of the Federal Council of 2018 (Bundesr?tliche Botschaft BBl 2018 591) explicitly state that the exception of Art. 24d also extends to commercial scientific research (pages 603 and 629 of the Comments), and they mention Data Mining, so that might apply here. But to the best of my knowledge, it remains to be decided by the courts where research ends and development of a product begins. Hence, it is presently hard to predict if the training of a machine such as Stable Diffusion falls under the exception of Art. 24d. (See also OFK/URG-Rehbinder/Haas/Uhlig, URG Art. 24d N6, where the authors find that the training of AI systems for commercial purposes is found to be outside the scope of Art. 24d.)

(Note that the EU Directive 2019/790 on copyright and related rights in the Digital Single Market is more precise here because it provides two exceptions in Art. 3 (for the purposes of scientific research) and Art. 4 (for the purposes of text and data mining), with the training of a machine such as Stable Diffusion clearly falling under Art. 4.)

Copyright and Derivative Work

Now let’s look at the second question. What if you (as a user of one of these systems) generate an image that comes close to a copyrighted original?

For example, this is what I got when I asked Stable Diffusion to give me Mona Lisa with a face mask:

As you can see, the original painting is clearly recognizable, but the lady is now ready to face a pandemic world.

Copyright sets a limit to deriving new works from original works. Art. 11 Paragraph 1 of the Swiss Copyright Law states that

1 The author has the exclusive right to decide:

a.?????whether, when, and how the work may be altered;

b.?????whether, when, and how the work may be used to create a derivative work or may be included in a collected work.

In simpler words, I cannot take an existing work, tweak it, and publish it.

In Mona Lisa’s case, this is not a major concern because the copyright of the original painting has expired. (Well, there may still be issues with heirs applying moral rights in certain cases and certain countries, but I won’t go into this now.)

However, when I asked Stable Diffusion to give me a Minion with a face mask, I got one of the yellow critters that again came very close to the original, and the copyright for the Minions hasn’t expired yet.

For that reason, I am unable to post that Minion-cum-mask image here.

Conclusions

As is apparent from the above, the training of Ais with copyrighted material may (under Swiss law) be a problem. Legislation in Switzerland isn’t very clear, alas.

More importantly, though, you (as a user of these systems) should be careful when using machine-generated images for commercial purposes. Such images may be very close to a copyrighted original, and you may not even be aware of that.

As to the second point, here’s some practical advice if you want to use AI-generated images for commercial purposes:

• Use a tool that allows you to provide your own artwork as input instead of using images from all over the internet. For example, the banner image of this article shows our tomcat Elmo combined with two different secondary images using Pikazo. All artwork is based on photographs I took myself.
• If you want to use Stable Diffusion, you can tell it to start from your own image as a base. It will still combine that base image with elements derived from other images, but the risk of generating derived work is smaller (see Stable Diffusion’s --init-img argument for details, but this will typically require installing an instance of the software on your computer, which is not for the faint of heart)
• Run your output image through an image search, checking if it resembles known, potentially copyrighted work. If you are using Stable Diffusion, you can compare your output image with the LAION-5B dataset, e.g. using https://rom1504.github.io/clip-retrieval/.

(Oh, and by the way, in most countries, the images generated by AI systems are unlikely to enjoy copyright protection per se unless they are based a creative, non-trivial prompting. For example, see Bundesr?tliche Botschaft BBl 2018 591, p. 620; or W. Straub in Schweizer IP-Handbuch, URG §12, I.A.4.3. Similarly, the U.S. Copyright Office will not register works created by machines. However, and once again, things are far from clear, in particular when the resulting work is a combination of human and machine effort. But this is not the topic of the present article.)

?

In conclusion: take care, or Stable Diffusion may come back to bite you.

?