You mean machine learning algorithms (or just "AI"), not Large Language Models. LLMs are just advanced word prediction machines; they're categorically incapable of detecting cheating in a game.

But, yeah. It would totally make sense to have server-side detection for things like:

• Consistency with performance (no changes over time within or sessions with reaction time, for example)
• Behaviour changes depending whether there is vs. isn't someone around a corner, particularly in areas with long sight lines so footstep sounds wouldn't trigger
• Inhumanly quick reaction times, with consistency (e.g. < 100ms reaction times — 101 ms is the world record)

etc.

Sure, people could still have cheats help tweak inputs, a bit, like "gentle" headshot aiming assistance, but it would catch egregious cheaters.