Every major publisher is talking about AI NPCs now. Characters that can hold dynamic conversations, respond to context, remember what you said three hours ago. It sounds incredible on paper. In practice, I’m not convinced it’s what games actually need.

Let me explain before the AI hype train runs me over.

What’s actually happening

Several studios have demonstrated AI-driven NPC systems in 2025. Ubisoft showed a tech demo where NPCs in a market scene could answer questions about their fictional lives, react to the player’s reputation, and generate contextually appropriate dialogue on the fly. NVIDIA’s ACE technology has been integrated into a handful of indie projects. Inworld AI has partnerships with multiple major studios.

The pitch is consistent: instead of pre-written dialogue trees with limited options, NPCs can have free-form conversations. You can ask them anything. They’ll respond naturally. The world feels more alive.

Australian studios are paying attention too. Several local developers I’ve spoken with are experimenting with large language model integration, though none have shipped anything to consumers yet. The interest is real, especially among studios working on narrative-heavy or open-world games.

Studios working with firms like AI consultants Sydney are exploring how these systems can be built efficiently enough for production use, not just tech demos.

The tech demo problem

Here’s my issue. Every AI NPC demo I’ve seen works great for about five minutes. You walk up to a character, have a surprisingly natural conversation, and think “wow, the future is here.” Then you keep talking, and the cracks appear.

The NPC starts repeating itself. It says something that contradicts the game’s lore. It responds to a serious question with something weirdly cheerful. It has no real emotional range — everything is delivered at the same moderate, helpful-assistant tone.

This isn’t a minor problem. It’s a fundamental one. Good NPCs in games aren’t just characters who can respond to arbitrary questions. They’re characters with personality, emotional arcs, secrets, biases, and specific ways of talking. A bartender in a Bethesda game doesn’t need to answer questions about quantum physics. She needs to feel like a real bartender in that specific world.

Pre-written dialogue, done well, achieves this. AI-generated dialogue, at its current state, mostly doesn’t.

What actually makes NPCs feel real

Think about the best NPCs in gaming history. Garrus from Mass Effect. The Companion Cube. Elizabeth from BioShock Infinite. Papyrus from Undertale. None of them needed dynamic AI dialogue. What made them work was writing, animation, voice acting, and carefully designed interactions.

The magic of a great NPC isn’t that they can respond to anything you say. It’s that every response they give feels intentional. A writer chose those words for a reason. A voice actor delivered them with a specific emotion. An animator made sure the body language matched.

AI-generated responses are, by definition, not intentional. They’re probabilistic. They’re “the most likely next word” repeated until a sentence forms. That works well enough for a chatbot. I’m not sure it works for a character you’re supposed to care about.

Where AI NPCs might actually help

I’m not saying the technology is useless. There are applications where it makes sense.

Background NPCs. The shopkeeper you visit once. The guard who gives directions. Characters who currently have three lines of dialogue and repeat them forever. Making these characters slightly more dynamic would improve immersion without needing the emotional depth of a main character.

Open-world ambience. In a game like GTA or Cyberpunk, having pedestrians who can react to context-specific situations more naturally would be nice. Not essential, but nice.

Accessibility. AI NPCs that can adapt their communication style could be useful for players with different needs. This is an underexplored application.

The cost question

There’s also a practical problem. Running LLM inference in real-time for every NPC conversation is expensive. Either you need server-side processing (which means always-online requirements and ongoing costs) or you need local inference (which means significant hardware requirements).

For a AAA game with thousands of NPCs across hundreds of hours of gameplay, the compute costs are not trivial. And someone has to pay for them — either the publisher eats the cost or the player pays more.

My prediction

AI NPCs will become a standard feature in games within three to five years. But they’ll mostly be used for background characters and optional interactions. The main story characters — the ones you actually care about — will still be written by humans, performed by actors, and designed with intention.

And honestly? That’s probably the right outcome. Not everything needs to be AI-generated. Sometimes the best technology is a good writer with a deadline.