Part of the big debate over generative AI is the question of whether it can be of any use in game development. Companies like Nvidia have made big claims about how it could fundamentally change gaming, but a lot of people seem skeptical. I’m skeptical too, but I think I’ve seen a path where generative AI might be helpful for specific kinds of games.

The first major controversy seems to surround the ethics of using generative AI graphics to make art assets for games. The data sets AI models use are so large that it’s impossible to confirm that they don’t include copyrighted or harmful content. Some of the people behind generative AI development have also either admitted to or been caught playing fast and loose with copyright. That might kill the technology for a lot of people right there.

Furthermore, you’ve got the line between a skilled artist possibly using AI to make some parts of their workflow easier, and people almost solely relying on AI to arrive at visual conclusions, which at best might just be recycling what already exists. Then you’ve got the wild claims about how AI might autonomously generate entire moving and interactive pieces of media.

Valve’s recently published guidelines for utilizing AI in games released on Steam mandates that developers disclose those techniques as one of two possible implementations. The other regards the use of AI models in-game, which sounds like completely untested territory.

Nvidia has made a big deal out of the possibilities for AI-generated non-player characters or even characters that dynamically generate responses to players’ words and actions. When you think about it, generating dialogue through AI is kind of like using AI to write a book. Yes, games are interactive, but many of them still tell authored stories. You can still tell the difference between an Elder Scrolls quest that Bethesda wrote by hand and the busywork that the game builds automatically and repeats endlessly. Still, that might just be the case for games designed around pre-written stories.

What made me seriously consider a path for generative AI in gaming was a ChatGPT mod somebody made for Mount & Blade II: Bannerlord sometime ago. If you watch the video, the experience looks extremely stilted and might reinforce any misgivings you have about AI. But if you understand how Mount & Blade works, the fundamental theory might have some merit.

I went over this after trying the original game for a while in 2019, but, basically, Mount & Blade is almost completely unscripted and populated by characters running on something that tries to feel like dynamic AI. Characters and factions constantly decide autonomously who to trade or go to war with. NPC dialogue directly reflects recent AI-driven events and characters’ immediate surroundings, and virtually none of it is written or directed by human developers.

If you ask a character for another character’s location, they might give you some guidance based on whether the target was recently seen nearby, what direction they were headed in, and when. The problem is that all of the lines sound extremely robotic — you can see the words and phrases that are being pulled out of a cloud and bolted together in each answer.

The ChatGPT mod makes all of that sound a lot more elaborate and less stilted. It also lets NPCs respond to player-written input, which could be a cool revival of systems from early text adventure games if it can ever feel smooth. If AI can make these interactions more dynamic, more fluid, and less artificial for situations that already have minimal human influence, then maybe, just maybe, it could do something interesting.

Looking at the mod also made me think about Arma 3, another game with significant amounts of completely autonomous gameplay content and systems, including dialogue. If you haven’t played it, squadmates in Arma relay information based on their actual surroundings. They’ll describe targets in terms of type, armaments, direction, and distance. Characters have a relatively large pool of inputs and responses to create a fluid flow of information in the middle of battles that never happen the same way twice.

What if generative AI could make characters more accurately and vividly describe what they see and where they see it. What if those characters could understand and respond to equally vivid information from players. It’s just something to think about.

It may turn out that in-game generative AI is of little use for many or most existing game genres. Maybe human-designed and human-written characters will always be better, but video games are a medium that sometimes involves significant amounts of automation. Like with much in gaming, maybe someone will figure something out that nobody can describe right now.