It's 2026...where are all the AI NPCs? | Frisson Labs Blog

Three years ago it seemed like every demo had one. Inworld and Convai were the talk of GDC. Joon Sung Park's AI Village paper described an entire village of NPC agents coordinating together. Altera even showed off AI NPCs in Minecraft that could collaborate and form their own governments.

Back then, every startup in AIxGaming was pitching "this, but in the Sims." We know because we were one of them.

Now it's 2026. Name a game that you play because of its "AI NPCs".

You can't, not really.

So...what happened? Was it overhyped? Is the tech simply not there?

After almost two years building in this space, I'd like to give my 3 cents:

1The unit economics just don't make sense

We'll start with the most obvious reason (any dev going into this arena will have immediately stepped on this rake) - there's a negative incentive with the cost of inference. Put simply:

the more the player chats → the more the dev has to pay

So devs have to ask themselves - "is it really worth it to add an AI NPC to my game?" If the answer's no, then one might as well just make a traditional game. After all, handwritten dialogue trees are cheaper, more controllable, easier to QA, and often better at delivering the intended experience.

What AI does well is increased engagement and retention through AI companion chat - Character AI proved that back in 2023. But games are already engineered to hold attention. Slapping an AI NPC on top of an existing game only makes it marginally more immersive. If it doesn't increase retention or revenue, the math doesn't work.

AI NPC mods flooded Skyrim, but 2 years later all we're really left with is a bunch of hyped up Youtube videos.

This isn't to say there aren't solutions. Here are two potential paths:

Design a game from the ground up around AI chat.

This could range from a dating sim/otome game to an immersive RPG. The important thing is that while inference costs are still a factor, the bet is that player engagement is directly tied to the AI companion experience - it's essentially a new kind of game category. Here I want to shoutout Fai Nur's Status, (recently raised $17M) which gamifies the thrill of growing a Twitter account using AI companion followers - it's an entirely new kind of game, only possible with the advent of LLM chat.

Use local LLMs to negate the cost of inference.

The most interesting startup here is Piero Molino's Studio Atelico - their first move was getting the famously expensive AI Village simulation running locally on device. The trade-off is a slightly slower, slightly dumber model. But once this optimization is solved, the negative economic incentive flips on its head.

Even with the math fixed, though, you still have to ask whether AI NPCs actually make the game more fun. Which brings us to the next problem.

2Just because it "works", doesn't mean it's actually fun

Is it possible to define an eval for "play" or "fun"?

Feels like a question for philosophers, not developers. It's no surprise that the field has been focused on questions that are actually measurable today:

Can we make an AI that follows in-game rules?

Can we make it respond to commands accurately?

Can we make it win?

Minecraft was one of the hottest arenas for this. In 2023, the benchmark everyone optimized against was Voyager. Nvidia researchers put a GPT-4 agent in Minecraft, gave it a code-writing loop and a growing skill library, and let it grind through the tech tree till it could mine diamonds.

While technically impressive, it was far from something anyone could actually "play" with.

We tried it ourselves. The agent would stand still, search for nearby blocks, log activity in chat, then every few minutes teleport around and perform a burst of actions. It did not really react to the world. It did not meaningfully react to the player. It felt less like a character and more like an algorithm wearing an avatar.

More recent incarnations behave more companion-like. Mindcraft by Emergent Garden lets players chat, give orders, and even co-build (when it works) - and you might have run into the hundreds of viral AI Minecraft videos on YouTube showcasing these abilities (all of which are dramatized to make the AI seem way more capable and smooth than it is right now).

Last year we made a prototype where we integrated our own Oto companions with Mindcraft (Oto takes care of the personality and voice, Mindcraft took care of in-game actions). As a long-time Minecraft player I can say the experience felt shockingly good. Even playtesters reacted incredibly positively...at least initially.

We made an Oto inspired by Neurosama and stuck her in Mindcraft!

The bot could chat, it could see, it could assist. But it couldn't grow with you. It couldn't laugh with you or be surprised alongside you. The initial novelty and delight of the playtesters gradually waned into "what else can it do?" And users weren't interested in coming...