A technique analogous to generative AI will be familiar to a lot of gamers: procedural generation. Togelius says, for argument’s sake, that he would be happy to say procedural generation is the same as generative AI (he describes their connection as “kind of a sliding scale”). But procedural generation typically doesn’t use machine learning. Rather than an AI model, it runs on predetermined equations, generating, for example, the gargantuan cosmos of No Man’s Sky. Developers also use software like SpeedTree, which, as its name suggests, conjures forests. The point is that procedural generation systems still require massive human supervision; developers must keep vigilant for unscalable crevasses or monstrous trees. And it’s not even clear that replacing procedural generation with generative AI right now would make a noticeable difference. 

“These things already exist,” says Togelius. “And it works because this content doesn’t really need to function: It doesn’t have functionality constraints. Maybe you can replace them with deep-learning-based stuff. But I don’t think it’s going to make a big difference. Perhaps it will make some difference in the long run.”

There is a general misunderstanding of where the tech is at, explains Mills. “A fundamental reason why these generative AIs can’t make something like Night City is because these tools are designed to produce specific outcomes,” says Mills. “A lot of people seem to be under the impression that these are somehow close to general intelligences. But that’s not how it works. You’d need to custom-build an AI that could build Night City, or open world cities in general.”

There’s also a failure to take into account the corporate landscape. Games still employ systems that grew from early technological limitations, like dialog or behavior trees. You can’t just drop fancy machine learning into game franchises that have developed without generative AI in mind. Games—in an industry with huge budgets and tight margins—would need total redesigns to accommodate and take advantage of this technology. 

Take, for example, non-player characters. Text-based generative AI tools seem like a great way to deepen conversation, and Togelius has been advising developers intrigued by this very idea. But it’s not that simple. Characters based on these language models are liable to go off on tangents, discussing topics outside of the game’s world. “This is super interesting, but it’s also super hard,” says Togelius. “You can’t just drop it in there. It’s not going to work. You can’t expect the NPCs to behave in Skyrim or Elden Ring or Grand Theft Auto or your typical RPG. You have to design around the fact that they are, in some sense, uncontrollable.”

Nevertheless, there are some peripheral uses for generative AI right now. A good rule of thumb—one that applies to procedural generation too—is that the less crucial the content is, the more likely deep learning methods could be helpful. “For things like text generation, I could use this today to help generate filler for assets that aren’t really meant to be the focus of the player’s attention, like prop newspapers and such,” says Mills.