I think AI searches could be much better even without an improvement in the technology. Probably the best AI patent search tool would be written by a ML engineer who has spent time examining a wide variety of patents. The current tools seem to be written by people who have only a cursory understanding of how patent search works, and that limits the usefulness.
For example, the current AI search tools don't seem to look at patent drawings at all. This is despite the fact that in many technologies, the drawings are the easiest way to determine similarity of the technologies. The words used vary a lot, but the drawings are frequently quite similar. Existing technology could be used to make a big improvement here, I think, but the problem is that people writing AI search tools seem to go for the easiest approach and only look at the text.
It does get more complicated than that. When I examine applications with flow or electrical circuits, frequently I'll run into circuits which are equivalent in some sense but arranged differently. An AI patent search tool should be able to handle this problem.
(Again, like my other comments here, this is just my opinion, not that of the USPTO or US government.)
