The real technical problem is making an AI that is working with limited information and trying to outsmart the player in a way that looks convincing, and that's mostly a strategy game problem. It's trivial to make deadly AI in an FPS because that's just a matter of tuning their aim cones and reaction times, and even basic squad tactics can be faked convincingly by just having them rotate out between "firing from cover" and "running to cover that's closer to the player or off to the side" in a staggered way so that half are shooting and half are advancing at any given time.

Strategy games are trickier and historically just cheat with perfect knowledge of the gamestate and tons of free resources, and they still get dunked on by skilled players. That is the genre that could benefit the most from machine learning on the task of inferring the gamestate through limited information and setting plans in reaction to known or guessed information. That's also been done by research teams, mind, but is probably out of reach of end-user facing systems for a few more years at least. Probably not until the next console gen at the earliest, if we're being honest. Not because the hardware required to run it locally is prohibitively expensive, but just because it's niche and probably more expensive than a majority of would-be consumers would be willing to shell out for.