Seems plausible. Reminds me of an article from long ago where a person used a training algorithm to get an FPGA to produce the behavior he wanted. The upside being that he got the behavior he wanted using less of the FPGAs capacity then a normally designed circuit would require. The downside was that it wasn’t reproducible on other FPGA chips. Whatever made it work required the subtle unique variations of that specific piece of silicon.

I have some vague recollection of this article. If I recall correctly the generated circuit consisted of two (or more) subcircuits that were not physically connected but managed to communicate using induced radio waves (which is definitely not something you would typically do with an FPGA).

Updated my comment with link.