The end of the dialogue tree: designing NPCs that actually listen

Think of the most memorable character you have met in a game. Chances are you remember what they said, not what you said back. For most of gaming history, your side of the conversation has been a short list of options written by someone who had to predict, in advance, everything you might want to say.

Language models change the shape of that exchange. A character can now hear what you actually typed or said, and answer in their own voice. It is the most visible shift in interactive storytelling in a generation, and also the easiest one to get wrong.

Talking is not the same as being someone

The first generation of “AI NPCs” proved that a model can chat. It did not prove that a character can exist. Ask a typical demo character about their childhood and they will improvise one. Ask again tomorrow and they will improvise a different one. They are fluent, agreeable and hollow.

An intelligent NPC needs more than fluency. In our work it comes down to four things:

• Identity. A stable core of who they are: their history, voice, values and the things they would never do. This is authored, not generated.
• Memory. What they know about the world and, crucially, about you. What you promised, what you broke, the name you gave them.
• Wants. Something they are trying to achieve that is not “help the player”. A character with their own goal will sometimes refuse you, and that refusal is what makes them feel real.
• Limits. Clear rules on what they know and what they can reveal. A dockworker should not explain quantum computing, and a villain should not hand over the plot because you asked politely.

A character becomes believable the first time they say no for a reason of their own.

Relationships as a game system

Once characters remember, relationships become a mechanic. Trust can be earned and spent. Rumours travel. A favour asked in one district changes how you are greeted in another.

This is where intelligent NPCs stop being a novelty and start being design. The model performs the conversation, but the relationship itself is tracked as state: numbers and flags a designer can see, balance and build missions around. The magic in the moment comes from the model; the meaning over time comes from the system.

The cast, not the character

A single clever NPC is a demo. A city of them is a game. Scale introduces its own problems: characters who all sound the same, who all know everything, who all agree with the player. The fix is the same one novelists use. Give each character a distinct voice, a distinct slice of knowledge and a distinct agenda, and let the friction between them create story.

In our own projects we write characters as short, dense profiles: age, background, temperament, what they want, who they distrust, how they talk. Those profiles are canon. Everything a character says is generated, but it is generated from that profile and checked against it.

Safety is characterisation

An open conversation can go anywhere, so an intelligent NPC must be designed to stay in role. We have found it most effective to treat safety as part of the character rather than a filter bolted on top. A character who knows who they are can deflect an inappropriate request the way a real person would, in their own voice, without breaking the world.

Where this goes

The near future is not one perfect AI companion. It is hundreds of modest, specific, memorable people living in a world that remembers what you did. Some will become your friends. Some will lie to you. A few will go on missions with you and talk about it afterwards.

That is the promise worth building towards: not characters that can say anything, but characters worth listening to.