Who if anyone owns copyright of algorithmically produced works?

Question

The image below is generated in real time by a Generative Adversarial Network trained on existing works of art (try reloading the page). The process is described in their paper which also demonstrates that it is indistinguishable by humans from art generated by contemporary artists and shown in top art fairs. They make no claim of copyright, freely provide the code that produces it and similar techniques have been used to produce many forms of work that would be copyrightable if produced by a human.

Can we say if it would be possible to successfully claim/defend copyright on such work in any way? Has anyone actually managed to defend such a claim? I am aware of the different decisions that have been made around AI as an inventor in patent law, but copyright seems very different. This question is somewhat prompted by this answer, but the question is different. Any jurisdiction would be interesting.

Answer 1

germany I am reasonably sure that works like this would be Public Domain, i.e. not copyrighted in the first place. In Germany, copyright (or more precisely, Author's Right) can only be held by a natural person, which is a legal term of art that essentially means "human being". There are two groups of natural persons involved here: the programmers and the creators of the works which are part of the training set. However, I can't see how either group could have a reasonable claim of copyright.

Note: there have been similar discussions about GitHub Copilot, an AI that writes code for you. It is my understanding that the process used by GitHub Copilot is roughly comparable to the process used here. With GitHub Copilot, there have been instances where significant snippets which are part of the training set have appeared in the output with only minor alterations. In this case, it is likely that the original author will have copyright over that portion of the output.

However, IFF a human being were actually involved in selecting specific works, then there is probably a copyright based on the creative decision of selecting this particular work and rejecting all the others. This is similar to the classic textbook example of a driftwood sculpture: the creative process here is not creating the sculpture but choosing to pick up this particular piece of driftwood instead of the hundreds of others on the beach.

So, if you simply generate these works and publish all of them, there is no copyright. If you generate a large batch and then select a certain number, then the person who made the selection might hold a copyright.