Whether that set is actually useful is a separate issue but someone is trying this over there for sure.
Yes, that's how games like Concord get made. Very successful approach to create art based on data about what's popular and focus groups.
I think what the previous comment meant was that there is data on how player play, and that tends to be varied but more predictable.
AI/fuzzers can't get far enough in games, yet, without a lot of help. But I think that's because we don't have models really well suited for them.
Edit: yup, it shut down nearly a year ago
Compare that to Helldivers 2 (online-only live service game, same platforms and publisher) which had a lot of personality (the heavy Starship Troopers movie vibe) and some unique gameplay elements like the strategems.
And sometimes it works; Apex Legends came out of nowhere and became one of the big live service titles. Fortnite did a battle royale mode out of nowhere and became huge.
Everything is measured and analysed and optimised for engagement and monetisation.
When you have 200 people making a game, "luck" or "art" doesn't factor in at all. You test, get data, and make decisions based on the data, not feelings.
Solo devs can still make artsy games and stumble upon success.
Where we used AI (machine learning, not LLM) was in terms trying to figure out what kind of human you would want to play with. We also used machine learning to try figure out what cohort of players you were in so we could tweak engagement.
Where LLMs could really shine, in my opinion: Gamers love to play people, not AI (now). People are unpredictable, they communicate, they play well but in ways a human could (like they don't have superhuman reflexes or speed). You can play all kinds of games against AI (StarCraft, Civilization, training of all kinds of FPS) but it isn't fun for long because you see the robotic patterns. However, an LLM might be able to mix it up like humans, talk to you, and you could probably make it have imperfect reaction time, coordination, etc. That would really help a lot of games that have lulls in human player activity, or too much toxicity.
I would be shocked if some games aren't doing this now. It seems like it still be hard to make a bot seem human, and it probably only works if you sprinkle it in.
Edit : oh yeah. A quick google search proved it : https://marvelsnapzone.com/bots/
I don't think it's much of a stretch to take this data over multiple games, versions, and genres, and train a model to take in a set of mechanics, stats, or even video and audio to rate the different aspects of a game prototype.
I wouldn't even be surprised if I heard this is already being done somewhere.