It's a fine line. There are plenty of stories of games that dumbed down their AI because it started doing inscrutable things when fully unleashed - and simulating full sense perception for something like a soldier in a shooter game is very computationally expensive, so "smart" opponents tend to feel like cheaters since they are thinking in terms of the underlying representations instead of behaving or reacting in a believable way.
A lot of making the experience of a game is in managing perception and belief - in making the player agree that the game is doing "what I expected" - and when a feeling of realism is aimed for, the amount of falsified behavior needed to get that result goes up dramatically. The games where belief is easy to achieve are mostly stuff where the ruleset is abstracted, empirically tested and legible to all - board games, physics games, puzzle games, etc. Things where the decision making itself is the intelligence. In contrast making soldiers that are satisfying opponents requires a blend of assets: carefully managing their positioning, animation states and voice callouts so that their decision making is something that players can identify, respond to and overcome.
"and simulating full sense perception for something like a soldier in a shooter game is very computationally expensive, so "smart" opponents tend to feel like cheaters since they are thinking in terms of the underlying representations instead of behaving or reacting in a believable way."
I am not saying it is easy. But today, we got some more cpu power to throw at the problem. And even if you simplify a lot, you would still get better results, than plain cheating, which breaks immersion for me too, when they all storm my position, they could not know exactly.
"In contrast making soldiers that are satisfying opponents requires a blend of assets: carefully managing their positioning, animation states and voice callouts"
And here is the crux: there are tons of ressources thrown at the animation states and voice callouts, which is even more frustrating, when their behavior is as dumb as allways.
And not everything is expensive to simulate: in allmost every game with stealth elements for example: you can kill some enemies and the others are looking for you for some time - and then they go back to their default state with "ah, it must have been nothing", even though they might have been already wounded by you and witnessed comrades fallen.
Here you would not need more cpu power, but simply a third state: where they can still give up the search, but remain on guard. (And tend to their wounds, etc. )
I don't understand how you can use stealth as an example. The only way you can fairly simulate stealth mechanics is by making every enemy render the scene at low complexity and then check if it can see pixels of the player model. Maybe some games do this but from my experience what happens when you get caught is pretty boring. They sound the alarm and they all hunt you down as if they had a radar because actually doing the simulation is too expensive. There is no need for a third state. Once the alarm is on it is on but it shouldn't mean the enemy knows where you are.
A lot of making the experience of a game is in managing perception and belief - in making the player agree that the game is doing "what I expected" - and when a feeling of realism is aimed for, the amount of falsified behavior needed to get that result goes up dramatically. The games where belief is easy to achieve are mostly stuff where the ruleset is abstracted, empirically tested and legible to all - board games, physics games, puzzle games, etc. Things where the decision making itself is the intelligence. In contrast making soldiers that are satisfying opponents requires a blend of assets: carefully managing their positioning, animation states and voice callouts so that their decision making is something that players can identify, respond to and overcome.