I think there's still a middle ground of realistically difficult AI that, like KobaCumTribute mentions, operates on limited information, and operates with lifelike reaction speed. You can easily make enemies have perfect aim, or more health, or whatever. But it's difficult to make them do that if you put human constraints on them. Like giving them slower reaction speed, or making them have imperfect "sight", like games usually do.

But I suppose that's not something you need machine learning for. Machine learning is only useful if it's used for actually learning how a player acts, or how players in general act. Because you can easily find the weak spot in the enemy AI and enemies become boring.

Game AI probably just needs more investment. I don't know if any game studios actually invest much in that stuff beyond re-implementing existing techniques. But good AI behavior can really make or break a lot of games.