Players have been asking for the ability to filter out games made with Gen AI.
We've added an automatic tag on SteamDB based on the AI gen content disclosures on the store pages.
Honestly, I’d love that as well, but the problem is that you cannot connect GenAI generations to mechanics because they’re too fuzzy. The best way to use them atm is to use them only for fluff. For example to automatically generate the art for encounters, or the flavor text for card games etc. But even then, they tend to converge into generic boring slop. Still I think there’s some potential there for some creative roguelike devs to do GenAI fluff kinda OK.
I think you have to be clever with the usage of gen AI to get non-boring things, and just use it as one or multiple elements in a larger pipeline/computation graph. This is my intuition and not battle-tested.
Disagree. They can be connected to actual game mechanics. For instance, it’s quite easy to ask an LLM to output something in json format:
{"name":"The Master of Evil","hitpoints":205,"class":"vampire,"}
and so on. You might object that it could make mistakes here. Suppose the detectable error rate is 10% (I actually think it’s lower from what I’ve played around with.) Rerunning it in the case of a such an error (e.g. malformed json, invalid class name, hit points exceeds bounds, etc.) reduces to 1%, then 0.1% etc., and in the end there can be a non-AI fallback just for certainty. Admittedly, the errors are not i.i.d., but still it should be pretty low. Many traditional procgen techniques, such as map generation, also use rejection sampling in this way, with even larger rejection rates than 10%.
It’s easy to generate something as generic as that, not as easy to generate mechanics. And if you don’t generate mechanics then you’re only doing fluff like I said
Honestly, I’d love that as well, but the problem is that you cannot connect GenAI generations to mechanics because they’re too fuzzy. The best way to use them atm is to use them only for fluff. For example to automatically generate the art for encounters, or the flavor text for card games etc. But even then, they tend to converge into generic boring slop. Still I think there’s some potential there for some creative roguelike devs to do GenAI fluff kinda OK.
I think you have to be clever with the usage of gen AI to get non-boring things, and just use it as one or multiple elements in a larger pipeline/computation graph. This is my intuition and not battle-tested.
Disagree. They can be connected to actual game mechanics. For instance, it’s quite easy to ask an LLM to output something in json format:
{ "name": "The Master of Evil", "hitpoints": 205, "class": "vampire," }and so on. You might object that it could make mistakes here. Suppose the detectable error rate is 10% (I actually think it’s lower from what I’ve played around with.) Rerunning it in the case of a such an error (e.g. malformed json, invalid class name, hit points exceeds bounds, etc.) reduces to 1%, then 0.1% etc., and in the end there can be a non-AI fallback just for certainty. Admittedly, the errors are not i.i.d., but still it should be pretty low. Many traditional procgen techniques, such as map generation, also use rejection sampling in this way, with even larger rejection rates than 10%.
It’s easy to generate something as generic as that, not as easy to generate mechanics. And if you don’t generate mechanics then you’re only doing fluff like I said
ah, I misunderstood by what you meant by “generate mechanics.” My bad.