Why No AI Games?

I have a theory.

Frank Kelly Freas, The Gulf Between

Here’s a puzzle — back in 2021, when the new AI era was just kicking off, it seemed obvious (to me at least) that it would lead to new kinds of games and radically new forms of gameplay. But here we are, 5 years later, and we haven’t seen anything to speak of. Clearly, AI is having a huge impact on how games are developed (like it has for all software), but there haven’t been any significant new AI-based game experiences. What gives?

What’s Out There

Maybe I’m exaggerating. Maybe there are some cool AI-centric games that I’m overlooking? Let’s see…

• AI Dungeon. This is the big one, obviously, and the earliest. The first version was built on GPT-2 in December of 2019. It peaked in popularity in ‘21, and has gradually declined since then. In my view AI Dungeon was always just a lightweight wrapper around the raw text-prediction engine. Most of its appeal was based on the underlying novelty of text-generation per se. I don’t get the sense that “prompt an LLM to tell you an interactive story” is a promising new kind of game people are especially excited about.

To be honest, that’s the only one I had even heard about. But apparently there have been a couple of other semi-successful ones…

• Death by AI. A party game where an AI game master presents a scenario, players describe their character’s actions and the AI decides who survives. I guess this was a viral hit? Apparently it got funding from a16z and some other VCs. Maybe this is fun. It wasn’t fun when I thought of the same idea a few years ago and playtested a prototype of it, but maybe I was missing something.

• Suck Up! Players are vampires trying to convince AI-based characters to let them in. Another version of conversation as gameplay. Look, I’m not going to tell you that I also thought of this game years ago and tried it out and it wasn’t good, because I didn’t. But you know what large language models are like, imagine doing a “funny” improv game with one.

There have also been a handful of semi-viral demos of AI-powered game “engines”. There was one that was a sort of Drunk Minecraft and Google has one called Genie 3 which seems technically impressive but also seems kind of boring. Even if it is capable of generating a reasonable facsimile of a stable, coherent 3D space you can move through and interact with, there’s nothing particularly mind-blowing or novel or interesting about these spaces. They just look like clunkier, sloppier versions of run-of-the-mill video games.

Am I asking for too much here? I don’t think so. Video games’ main job is to blow people’s minds. They are extremely good at highlighting the things that are amazing about computers. I distinctly remember the shudder of sublime metaphysical weirdness that went through me the first time I played DOOM on a LAN and turned the corner and saw another person who was inside that imaginary space with me. I remember having my mind blown by MYST. I remember my first encounter with the linguistic magic of Infocom’s parser-based adventure games. I sometimes feel some of that magic when interacting with LLMs, but nothing remotely like that from any of these AI games.

Is There Also a Surprising Lack of “Good Old Fashioned AI” in Games?

Thinking about this topic made me realize that something similar was true of AI in games in general, separate from the question of these new models. There just aren’t all that many examples of games where interesting AI is the hook. There’s The Sims, of course, but is that even really AI? There are those Orcs in that one Lord of the Rings game who hold grudges against you. There’s the alien in Alien Isolation. The creature in The Last Guardian. The enemy soldiers in Fear? The creature in Black & White? The bickering couple in Façade? Am I missing any big ones? Doesn’t it seem like there should be more?

Let’s hold off on that question, and return to the main puzzle. Why hasn’t this miraculous new version of computational magic given us a single truly groundbreaking game?

Three Possible Explanations

The first two are prosaic, practical issues and the third is a “design theoretical” speculation that I think explains what’s really going on here.

1. Business Models. It’s very hard to build a real game around core functionality that you are paying a third party to supply. I’ve built prototypes that were reasonably fun but there was no way to actually release them that made sense. Charge players a subscription? Some kind of microtransactions? Ironically, when it first launched, Death by AI nearly went bankrupt due to OpenAI/ElevenLabs costs. This dynamic also discourages developers doing small experiments and releasing them for free, hoping to go viral. The incentives are all wrong. Developers are highly motivated to hit the model as little as possible, to use cached, pre-generated responses or find other workarounds. I’ve also built game prototypes where the whole experience changed dramatically, for the worse, because the model I was building around changed in ways I couldn’t understand or control.

2. Culture Wars. The video game audience has decided, almost unanimously, that they don’t want generative AI in their games. This might eventually change, but as of now it is a well-established, non-negotiable taboo.

3. The Nature of Fun. But I don’t think either of the first two reasons would be enough to prevent a truly compelling new game experience from making itself known. I think what’s really going on here is that these new models just aren’t the natural source of fun I thought they would be when I first encountered them. Some forms of computation end up being a deep well of fun. 3D rendering is fun. Physics engines are fun. But these new AI models, despite being powerful and useful and fascinating, don’t seem to be intrinsically fun. Which I think is kind of surprising!

The Nature of Fun

Years ago, when I first started playing around with the new models, I thought their messy, organic, stochastic qualities could be the key that unlocked whole new categories of gameplay. The story I told myself was that traditional video games were overly “mechanical” — they relied on brittle, deterministic logic and formulaic “lock-in-key” behavior. I thought that, if we embraced the psychedelic weirdness of these new models, we would discover a new kind of dream logic that would support wild new kinds of surreal game experiences. I think I was wrong.

The past few years of playing around with these things, which ranged from little experiments to full-scale commercial projects, has left me with a new-found admiration for the brittle, deterministic, mechanical logic of old-fashioned video games.

You know what’s fun? A stick. A stick is fun. A ball is fun. I now have a (hopefully) more nuanced appreciation for the way that the fun of games is rooted in simple behaviors and deterministic rules. The fun of games is deeply connected to the miracle of emergence, in the ways that a small set of seemingly trivial constraints interact with each other to produce an infinite expanse of surprising complexity. Starting with a bunch of surprising complexity doesn’t lead to even more fun, it just short-circuits the whole process.

The soft logic of generative AI is too much like the soft logic of other people. Other people aren’t intrinsically fun. They’re sacred and profound and fascinating and beautiful but they aren’t, by themselves, fun. That’s why we have games!

I guess I should have seen this coming. I knew the best way to understand these models was as simulators, and I may have fallen into a version of the immersive fallacy.

I’m sure that, even as I type these words, there is a clever teenager somewhere proving me wrong. I certainly hope so. There’s no way this crazy new form of computation won’t eventually lead to some wild new kinds of gameplay, and I fully intend to keep messing around with these things looking for it. Only hopefully now with less panicked urgency and a slightly sharper eye.