That is a wild extrapolation from a sample of 615 people
This survey was conducted online by The Harris Poll on behalf of Google Cloud from June 20, 2025 - July 9, 2025 among 615 adults aged 18+ working in game development in the United States, South Korea, Finland, Norway, and Sweden.
So, it’s a voluntary poll. That’s a great way to get a biased sample.
Also, some of the responses sound like Google is playing fast and loose with the term “AI.” Is procedural world generation AI? Google seems to think so, despite it existing long before LLMs we’re a thing.
This whole thing reads like “research” designed to promote AI. I wonder why Google might want that? /s
🙇♀️ Page 19 shows how biased their narrative is.
🚮Appropriate headline:
We made a narrative based on a short random survey in 5 countries based on vibes.
What percentage of games are shovelware and asset flips? Some “game developers” have always been pushing slop.
Is procedural world generation AI?
Its’ not if you don’t use AI for it. But, many many people wish they could use AI for worldgen to see what they can get out of it.
I feel like we have pretty good concepts and tools for world gen. If games are going to use AI I’d expect the best results to be in quest systems. Most games struggle to procedurally generate good quests. They end up feeling formulaic in almost every aspect.
I imagine even if the quest structure stays the same simple “go to X and do Y” but you let AI generate a good reason to go on the quest it instantly improves the quality of procedural quests.
But that assumes AI can generate a good reason and I have my doubts about that, and a lot of other things AI supposedly can do.
I only really enjoy WFC mapgens, simply because structures they come up with are largely driven by modular pieces and that produces interesting results for longer time (for me personally). I find noise/biome/temperature driven open-world worldgens boring af, and I get a feeling I’ve seen it all very very fast. AI can potentially produce unique structures in open-world worldgens way better than noise-based algorithms with basic heuristics on top of them. You mentioned quests, just consider that AI-based worldgen can generate/modify world based on those characters and quests. You can ask it to start with noises, then modify to arrange for villages/cities, then make sure there is nice road from village A to village B and landscape is modified to make this road nicely traversable, and if there is a quest, modify map in a way that the needed dungeon happens according to intended progression in the mountain between those villages, etc.
LLMs and other machine learning are just algorithms. That’s all procedural world generation is, and this insistence by the Tech Bros that we need their models to “boost creativity” is a farce.
My opinion? The people that wish they could use AI to “see what they get out of it” are lazy ass fucks who don’t want to put in the extra work to actually get good at game dev.
I want to see it myself real bad. The reason for this is actually very simple: more traditional handcoded worldgen algorithms usually operate with some basic noise functions controlled by some parameters like “biome” or “temperature” or “height” and then slap some heuristics on top to smooth rough edges or to introduce a bit more of interest. Those heuristics you code there are rather limited. You ofcourse could spend a lot of efforts and hardcode a lot of stuff there, but it’s still limited. And in practice they are most often are very limited. With AI though, what developers can hope for is multistep generation with self-feedback. We may manually model some prefabs, modular pieces and ask AI to stitch them together in a way that resembles some special symbol per map, possibly generating some intermediate pieces by itself if those are lacking, also come up with enemy placements and look at the thing at whole and try to rebalance it for certain difficulty, etc. It’s more flexible and it’s potentially unbounded. You can ask it to reprompt itself however times needed if it see there are some problematic places or missed opportunities in map it generated. You can give it a list of gimmicks and ask to try to compose and balance every map around random gimmick picked from this list. You can also ask it to roll a dice and with probability of 15% it will invent a gimmick itself instead of picking from the list. Possibilities are wild.
Nothing of what you suggested is particularly difficult with real dev work. You basically just said, “I want to vibe code it all.” It’s trivially easy to set up pseudorandom generators; deciding where enemies and objects go should not be left up to chance through some black-box algorithmic “magic.” Game theory exists for a reason, and AI doesn’t “know” about it, because it’s just a complex pattern generator at the end of the day.
Also, what happens when the model generates an environment that can’t be traversed? What if it places invisible walls in weird places? What about an environment that’s rife with bugs? What if the code is plain wrong? Now you have to go into the code, learn how it works, and debug it manually. Thank god you saved yourself some time by vibe coding. /s
I can see we won’t agree, so you’re welcome to get the last word, but I won’t reply afterwards.
Another bad faith / inexperienced take.
Also, what happens when the model generates an environment that can’t be traversed? What if it places invisible walls in weird places?
That’s also one of the reasons why it’s interesting. This happens a lot when implementing regular mapgen and you have to fix it until it only generates correct maps. AI can perceive what it generated and make sure certain invariants are holding and if not, modify map to fix it, and continue going and going. You can ask it to start with noise and carve space for villages and carve roads between them. You can ask to start with noise and quests and generate roads based on what makes sense for progression, and so on.
X for doubt
