Is an AI model considered a derivative work?

Question

Suppose an AI art model was trained exclusively on Mr. X's artwork. Suppose Mr. X managed to prove that the model was only capable of producing derivative works. He would then own the copyright to all the art produced by this model. But would Mr. X also own the AI model?

Now, let's take this question one step further. Suppose an AI art model was trained on art generated by millions of different artists. And suppose these artists managed to prove that the model was only capable of producing derivative works of at least one of these artists. Would the model then become public domain?

Whether AI-generated art is derivative has already been hotly debated, and that is not the point of my question. If the artwork is proven to be derivative, who owns the model?

Answer 1

There is no provision for automatically relicensing infringing works (for example, distributing a program that contains parts covered by the GNU GPL and that is therefore a derived work will not automatically place the infringing program under GPL, even if that is the expected way for derived works to comply with the license.

Instead, the derived work becomes at the very least undistributable, as there are competing copyright holders that disagree.

Answer 2

No one has copyright in the model

The person who created the code has copyright in the code. The person(s) who provided the training data, assuming it’s something subject to copyright (art is, facts aren’t) owns copyright in the input(s).

As an artefact of a computer algorithm, no one has copyright in the model. Similarly, no one owns copyright in the non-original parts of the output art although it can still be infringing.