I did a conferencr talk maybe three years ago on using AI for games.

Since the talk, code, audio,and imagery have progressed leaps and bounds, almost all of the points on active AI use haven't changed much.

You have more capable smaller models now that means you could at least run local models in games, but you still don't get to see what it is going to do before the user does. Developers are acutely aware of the climate. if a game has a fault because the AI does something unexpected, it's not just a bug, It's a news story.

It still cannot be realistically used for multiplayer games. The models are prone to adversarial attacks, which can harm a multiplayer experience.

I did point out at the time that solo sandbox games could benefit. An active AI can be empowering for the player, unchecked empowerment can be great in a solo game. It destroys a multiplayer game.

Then there's just a lack of training in some of the specific areas of games. You pretty much have to filter game state down to text for a llm, I don't yet know of any gameplay embedding model.

I have played around with a little tower defence to see how much an AI can work on gameplay, It essentially has a REPL interface to the game logic where it can take actions and advance the game a tick at a time. It does an ok job, but there's still not much understanding of time and urgency to work with an active environment