Who owns this AI-generated painting?

Do androids dream of electric creativity?

 Habitual readers may have noticed that I have been obsessed with artificial intelligence copyright for quite a while, almost as long as I have been interested in simian authorship. For anyone interested in revisiting some of the arguments, I have written an article on this topic. The subject has exploded in the last couple of days in my timeline with the news that a painting generated by an artificial intelligence program has been auctioned and sold for an astounding $432,500 USD.

There are quite a few interesting questions arising from this. Does the painting have copyright? If so, who owns it, the programmer or the person who pressed the button? Why would anyone pay such amount of money for that? Did Sotheby’s check the frame this time to avoid a Banksy?

Questions. Questions that need answering.

But first a bit of background. The painting is called Edmond de Belamy and it was made by the French art collective Obvious using a machine learning algorithm designed specifically to generate images, known as a Generative Adversarial Network (GAN). The artists fed the AI over 15,000 portraits from various epochs, and produced a set of portraits of the fictional Belamy family (the name is a play on words in French, comes from “bel ami”).

The Belamy family tree

So, does the work have copyright? As this was made in France, European copyright applies, and in my opinion almost certainly none of the paintings have copyright, and therefore would be in the public domain. Copyright subsists if a work is the author’s own intellectual creation expressing their creativity; given the process described by Obvious, there is definitely some input by the artists in the manner of the selection of paintings to be mined by the AI, as well as the selection of the resulting “family”. There is also some element of creativity in the reproduction of the artwork into an actual painting. On the whole, I don’t think that this would amount to enough creativity, but it will be interesting to see if a court agrees, we may have to hear evidence of the exact process in detail.

With regards to the actual physical paintings, they do not rely on copyright, which is why one of them has been auctioned and sold. An interesting question may arise as to whether people are able to reproduce the painting without permission, as it can be assumed that it’s in the public domain, it should not be an issue.

So far, so good. But there is a twist to the story. The GAN algorithm used by Obvious was not created by them, but it is a program created by researcher Ian Goodfellow, and it is an open source project released under a BSD licence. Under the terms of the licence there is nothing wrong here, developers are allowed to create derivatives from the software, and this includes different versions of the program. Obviously, there is no chance that Goodfellow would be able to claim any ownership of the resulting artwork. However, Robbie Barrat, a person working at a Stanford research lab, now claims on Twitter that he is the one that trained the AI one year ago, and has linked to his data on github. Barrat claims that he was asked by a member of Obvious for help on getting the models to work. Obvious responded that they did credit Barrat in their papers.

The ensuing discussion is fascinating, it always amuses me to see software developers tackle legal questions, some people are accusing the French artists of being “obvious-ly thieves”, while others seem to argue that this is an open source project, and therefore Obvious can easily do anything they want with the outputs.

The problem at the heart of the discussion is that people seem to confuse three different things: software, data, and the resulting artwork. Software is protected by copyright and it can be licensed, it seems like Barrat used GAN to train the AI, there are no issues there. Data may or may not be protected, as Barrat is in the US I would argue strongly that his outputs are not protected at all. Furthermore, the resulting images from the training of an AI are probably not protected anywhere but in a few jurisdictions (UK included, read my paper). Therefore, Barrat cannot make any ownership claims of the images as these are not creative works, and therefore not protected by copyright.

I go into a lot of detail about this in my paper, but long story short is that the US Copyright office has declared that only humans can create works worthy of protection, and this seems to be the case in many other jurisdictions. As of now, most AI works will be in the public domain and anyone can re-use them. Conflicts like the one described above are precisely what I argued would happen, as more and more AI works will have commercial value.

It will be fascinating to look at what happens, in my opinion, we have not heard the last from the Belamys.