Lately, I’ve been toying with the idea of using LLMs as NPC actors in a video game. It would be fascinating to watch them interact with each other and the world, creating their own emergent stories.
Ideally, the player wouldn’t directly interact with the NPCs in a first-person sense. Instead, you’d observe their autonomous behaviors from a distance — like watching a small simulated society evolve.
To make things more interesting, the player could influence these agents indirectly by implanting thoughts into their minds. Imagine a small village of NPCs going about their daily routines, and your goal is to shape their world. Maybe you convince them to worship the village cat as a god, or perhaps you sow chaos by influencing the baker to poison the bread.
Since these NPCs are LLM-driven, the range of emergent gameplay could be far broader than in traditional sandbox games. Crucially, not everything needs to be explicitly programmed — it’s enough if the NPCs believe something exists (like poison) and act accordingly.
Limitations and Potential Pitfalls
Using LLMs in this way has at least two major challenges:
Cost and Infrastructure
Running LLMs is computationally expensive.
Local execution is possible but tricky. Using APIs from providers like OpenAI, Google, or Anthropic is far easier but introduces cost concerns.
That said, API costs are dropping quickly, and hybrid strategies could mitigate expenses — for example:
- Dynamically switching between expensive and cheaper models depending on task complexity.
- Using a graceful degradation model, where high-end models are used until a quota is reached, then temporarily replaced with cheaper alternatives.
Unpredictability
The second issue is unpredictability — both a strength and a weakness.
It can produce surprising, emergent behavior and high replayability from a simple core loop. But it also makes it difficult to anticipate what players might do. Players may exploit or subvert the system in ways you never intended — which could be both delightful and chaotic.
The Pitch: Intrusive Thoughts
Intrusive Thoughts is a top-down 3D simulation game
(think: The Sims meets The Office)
where every NPC is controlled by an LLM.The setting is a mundane office environment — think Dunder Mifflin Paper Company.
Each character has a unique personality, backstory, and set of goals.As the player, your only power is to plant intrusive thoughts into their minds:
- “Eat your coworker’s lunch in front of them.”
- “Order a thousand rubber ducks and scatter them around the office.”
- “Start a petition to replace coffee with soup.”
- “Whisper to your coworkers that time has been looping since Monday.”
From these small seeds, entire stories unfold organically as the LLMs act, react, and improvise.
The initial prototype would stay deliberately small in scope — one office map, a handful of NPCs.
This keeps the interactions dense and focused, making it easier to test the emergent behavior loop.
World interactions should prioritize versatile mechanics (like “pick up” and “drop”) instead of specific, pre-scripted actions (“drink coffee”). The LLM can roleplay the details — e.g., writing *drinks coffee from mug* — without needing a custom animation.