Who owns AI-generated inventions?

An article about AI recently made me think about whether AI-generated IP is a real thing, and whether it would be an issue for IP law as it currently stands. The article concentrates mainly on copyright, and comes to the conclusion that a human would still be the author of a work generated using an AI system. I don't generally deal with copyright stuff, but that seems fairly uncontroversial to me. There is also a short section on patents, in which an interesting hypothetical example is given of a case where inventorship might be in question. Here's the example:

"Person A builds an AI system for identifying the best material to use given a series of input parameters relating to a particular engineering context. Person B takes the system and uses it to search for the best material to use in the design of a tail fin for an unmanned aerial drone. Person C then looks at the three potential options identified by the system and decides to file a patent covering the use of the third material for tail fins in unmanned aerial drones, as they recognise that it will have particular technical [benefits]."

The question is then who is the inventor? An answer is not given in the article, which simply states that a court or patent office would need to decide who was entitled to be named. This to me seems to miss an important point or two, as well as being a bit of a cop-out.

Firstly, let's say the invention is about the use of a specified material in a tail fin for an aircraft (why it has to be an unmanned drone is not clear to me, so let's leave that aside). The inventive concept, given that tail fins for aircraft are known, is the use of the specific material.

Person A has devised a system for identifying a material given a series of inputs. By itself this has nothing to do with the invention, so this person can be discounted as being the "actual deviser of the invention" (section 7(3), UK Patents Act).

Person B has plugged some numbers into the system created by person A, and the system has output some candidates. Person B has not arrived at the invention, but if they had the thought of using the AI system to search for candidate materials they might be considered to be a joint inventor. This is similar to the concept of the skilled person being different for devising the invention compared to understanding the invention, something the case of Schlumberger v Electromagnetic Geoservices was concerned with. It is not clear from the example, but if person B had the idea of using the system to search for a candidate material, this connection could form at least part of the invention in the same way that the idea of connecting two separate technical fields to arrive at a new combination could be. Person B might, however, have been acting on orders from someone else, in which case another person might get involved.

Person C is the clearest one for being an (if not the) actual deviser of the invention, since they come up with the material on which the claimed invention is based. They should be a named inventor.

The solution to the problem is actually not really anything to do with AI. That might be to do with the example being a poor one, but it might also get to the point of whether an AI system can ever be considered to be an inventor if it needs to be given an input to provide a desired output. If an AI system can truly think for itself, it would not necessarily be possible to trace the particular input that generated a particular output (it might have been lots of different inputs, as would be the case for a human inventor), so identifying an inventor could prove more problematic. At the moment, however, as far as I am aware such a system does not exist.

A final point about the AI system in the example is also worth considering. What the example does not say is whether the AI system is publicly known or is secret. If the system is already known, the technical problem to be solved becomes how to select a material for a tail fin, to which at least part of the solution becomes obvious: ask the computer. Whether it would be obvious to select one material from three is another question. In some cases it might be.

Perhaps the lesson at the moment is that we are not yet in a position to even think clearly about whether an AI system can really be the actual deviser of an invention, for the simple reason that AI does not really exist yet. If it ever does (and, if Roger Penrose is correct, it might never happen), so many things would change that the question of deciding which entity is entitled to a patent may become insignificant.

UPDATE 22 June: For further reading here is a recent journal paper that goes over the points raised here, along with many others, and in a lot more detail.