NPCs and AI: What Makes a Character Convincing?

Written dialogue gives an author control

A dialogue tree can offer carefully timed choices, performances and consequences. Limited choices can be deliberate: an author knows which information the player should hear and what the character is allowed to reveal. A fixed line is not evidence of a technically simple world.

A behavior system can separately decide where a character goes, which activity it performs or how it responds to a nearby event. Speaking and acting are different parts of the simulation. The fact that a character moves convincingly does not tell you how its dialogue was created.

Generated replies change the input, not the whole game

A conversational model can respond to free-form text or speech instead of only selecting a written line. That can make an interaction less predictable. The application still needs a character definition, a boundary for permitted topics, a connection to world state and rules for actions.

An invented promise is a useful example: a character might say it has opened a gate even when no authorized game action has occurred. Fluent text does not modify world state by itself. The integration must check the action and report the real result.

Sources: Convai · Mindview character layers

Memory is a designed feature

Conversation history, retrieved facts and saved character state are not the same thing. A system may retain recent messages while forgetting an earlier correction, or save a short summary across sessions rather than the entire exchange.

Inworld’s current memory documentation describes persistence as an application responsibility across sessions. A good demonstration should show which facts are saved, how they can be corrected or deleted, and what happens after a fresh session. “It remembers everything” is a claim that needs much stronger evidence.

Sources: Inworld · long-term memory and persistence

What research can illustrate

The 2023 Generative Agents paper studied a 25-agent sandbox combining observations, retrieval, reflection and planning. It is useful because the architecture connects memory and behavior rather than treating an isolated reply as the whole agent.

That research example is not a description of GTA VI. Neither a vendor’s character demo nor a research simulation establishes which technologies Rockstar uses. We have not verified LLMs, a particular memory system or autonomous agent behavior in GTA VI.

Sources: Park et al. · Generative Agents, 2023 research paper · Rockstar Games · Grand Theft Auto VI

Four checks for a character demonstration

• Dialogue: can it stay in character when asked a surprising question?
• Memory: does a corrected fact survive a new session, and can the user remove it?
• World rules: can the character perform only permitted actions, and does it distinguish a request from a completed action?
• Reliability: does it admit missing information, and is the visible behavior consistent with what the interface claims?

An illustration you can use to think about it

Imagine a station clerk. A written version offers three destinations. A conversational version accepts “somewhere quiet by the coast.” Both still need real destination data and a booking rule. If the model invents a place, the world integration must reject it rather than silently creating a fake ticket.

This is an editorial illustration, not a live model benchmark or GTA gameplay. The technological continuation on CheatCode explains current character tools and a reproducible way to assess their limitations without assuming they are features of a particular game.

Where to go next