Every AI Game Is a Chatbot

Every runtime LLM game is building the same thing. Talk to the NPC. Type at the character. Have a conversation. The demos look different but the interaction is always the same, and I think that's the mistake.

Frank Lantz asked Why No AI Games? ¹ recently and I mostly agree with him. The business model is bad, the culture war baggage is real, and the nature of fun is a hard problem. But his post helped me narrow something I'd been poking at for a while. We don't have the killer app yet. Doom proved the FPS. Wii Sports proved motion controls. Beat Saber proved VR was fun. Half-Life Alyx proved it could be immersive. Runtime LLM games don't have anything like that. We don't even have the small hits. And I think part of the reason is that everyone converged on chatbots.

Chatbots Everywhere

The most common runtime LLM game pitch is some version of "talk to the NPC." Maybe you type. Maybe you use your voice. Maybe the NPC remembers what you said. Maybe your squadmate talks back in natural language. It's obvious immediately, which is why it demos well, but typing to a character isn't a core pleasure of games.

Mario is fun because traversal is fun. The verb is good. You start there.

A lot of runtime LLM game work misses this. The technology is impressive, but the interaction is boring. It's not an interaction I want to repeat for hours the way I want to keep moving through Mario, reading lane pressure in Deadlock, pulling the camera from the campaign map down into a cavalry charge in Total War, or wandering endlessly in Minecraft.

This is also why I keep coming back to the Square Enix Portopia GDC talk from 2023. ² I wrote about it at the time because it felt like an early clear statement of the pattern. Let the player type freely. Use NLP or an LLM-like system to interpret the input. Route the result back into authored game logic. Fall back to chit-chat when the system can't cleanly map what the player said.

It was interesting because it wasn't saying "let the AI run the game." It was saying "use language models as interaction glue." The game still knew what it wanted. The system was just trying to meet the player halfway.

Three years later, I think that pattern has been proven out. I also think it has a ceiling.

Where Winds Meet

Where Winds Meet ³ is the best real deployment of that pattern I've seen so far. It's polished, it shipped, and it makes the case for a practical use of runtime LLMs better than most of the louder AI NPC demos.

The important thing is that it's not using LLMs as game logic. The quests, state changes, progression, and rewards are still authored. The LLM sits on top of that deterministic structure and handles player expression. You type naturally. The NPC responds in character. The system evaluates whether the conversation has meaningfully satisfied what that interaction was looking for.

Much more grounded than "the NPC can say anything."

It also helps explain why the system works at all. There are still traditional NPCs in the game. Then there are the semantic ones, where the conversation is doing fuzzy input work over an authored state machine. The game isn't asking for one exact line, but it's still looking for something. A kind of answer. A tone. A semantic shape.

Clever, shippable, still a chatbot.

And that's my issue with it. The system may be more sophisticated under the hood, but the player-facing interaction is still typing at a character until the game decides you've landed close enough to what it wants. It's open-ended in presentation, but not really open-ended in the way people mean when they fantasize about AI games. More importantly, it's still not a game interaction I find especially fun.

I don't mean that as a dunk on Where Winds Meet. I think it's the first real proof that this approach can survive contact with an actual shipped game. But it also makes the ceiling clear. If the best current version of runtime LLM interaction in games is still "chatbot, but better structured," then it shouldn't surprise anyone that nothing has broken through yet.

The Better Exceptions

The example I keep circling that feels most different is MIMESIS. ⁴ I haven't played it yet, so I don't want to overstate the case. But it reminds me of a Lethal Company mod I played that used voice cloning to make the monsters speak in other players' voices. That was genuinely unsettling in a way that felt like a new interaction, not a more elaborate conversation. MIMESIS seems to be chasing something similar, using voice models and LLMs together in a way that's actually searching for a new kind of play.

I think about Infinite Craft ⁵ in a similar way. I'd put it more in toy territory than game territory, but it's good. It matters because it suggests the interesting uses of these systems may look more combinatorial, more systemic, and less like talking to a person.

I want more of that. Not "what if the NPC could really talk," but "what kind of system becomes possible when language generation is embedded inside the rules of play?"

I don't have the answer to that. I do think the answer is probably not open-ended typing.

The Local Model Problem

Lantz's business model point is a big one. I agree with him. It's hard to imagine a lot of these ideas surviving remote inference costs, unpredictable latency, and the general ugliness of building a game around metered API calls.

Part of why I built RogueLLMania. Not because I think I found the answer, and definitely not because I think I made the killer app, but because I wanted to feel the constraints directly. Could I run a model locally? Could I use it in a tightly bounded way? Could it sit inside a deterministic game instead of replacing one?

My answer was basically: maybe, but the models are still weak. I deliberately tried to use a small dumb local model and couldn't get results I liked. Still, I came away thinking local is probably the only business model that makes sense for games. If this stuff ever really lands at runtime, I suspect it'll be because game developers did what they always do, made weak hardware sing through constraint and optimization.

We Are Not There Yet

I think Frank Lantz is right to ask why there are no AI games. I just don't think the current wave gives us enough evidence to close the case.

What we have instead is a field of experiments converging on the most obvious use of the technology. Chatbots. Voice NPCs. Free text input. That pattern can be polished. Where Winds Meet proves that. It can even be interesting in small doses. But I don't think it's the breakthrough.

Maybe that's because runtime LLMs really are a bad fit for games. Lantz puts it well: "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 take that argument seriously.

But I also think we've barely started exploring what these systems look like when they're constrained, embedded, and used for something more game-like than conversation.

We're still waiting on the Doom, the Wii Sports, the Half-Life Alyx. Maybe that moment never comes. I wouldn't rule that out. But I also don't think "we turned NPCs into chatbots and it was kind of boring" is enough to answer the question.

-b