We did this in a previous org. Basically, we had a bunch of user-generated data, users would then classify a sample of that data, and then we’d train our model on those classifications.
I don’t see how it would work in game dev though, unless they’re using AI to customize an NPC’s behavior based on the player’s actions (i.e. teaching an enemy to block player attacks). Generating models and whatnot would just have too small of a data set to work with.