The non-player character has, for most of the medium's history, been a thing that said what it was written to say. A dialogue tree offered the choices the player could make, and the character responded with one of the lines the writer anticipated. The character was predictable, repeatable, and confined to the script. What is arriving now, in a small but growing number of games, is a character that can converse — and that conversation is changing what a game character can be.

The shift is driven by the same language models that have reshaped the rest of the technology industry, applied to a problem games have had for decades: the brittleness of scripted dialogue. A character that can respond in natural language to what the player actually says, rather than to a pre-written option, is a character the player can talk to rather than read.

The coherence the conversation requires

The hardest part of a conversational character is not generating words; it is staying consistent. A written character is consistent because the writer ensures it is. A character driven by a language model has no such guarantee. It can forget what it said moments ago, contradict itself, or drift from the personality it was supposed to have. Each of these breaks the illusion that the player is talking to a person rather than a machine.

The more thoughtful implementations address this by giving the model a persistent description of the character it is supposed to be — a background, a set of beliefs, a way of speaking — and constraining the conversation to remain within it. This is not a solved problem, but it is a manageable one, and the characters that result are coherent enough to sustain a conversation that a player can take seriously.

The safety the conversation demands

A character that can say anything can also say things it should not. The same models that produce a believable conversation can produce harmful, off-brand, or simply wrong output, and a game that lets a character converse with the public has to ensure that conversation stays within bounds. The safety work is less visible than the coherence work, but it is no less critical — a single widely-shared clip of a character saying something damaging can define a game's reputation.

The implementations that take this seriously build explicit boundaries into the model's instructions, screen output against safety filters, and log conversations for review. The ones that do not are betting that the model will behave, and that bet has repeatedly proven unwise. The technology is now mature enough that the absence of safety engineering is a choice, not an oversight.

What the character becomes

The deeper question is what a conversational character means for the medium. A written character is an artifact; a conversational one is, in a sense, a participant. The player's relationship to a character that can respond to them is different from the player's relationship to one that cannot, and the design implications of that difference are still being worked out. Some genres — mystery, social simulation, role-playing — are natural homes for characters that can be interrogated or befriended. Others gain little.

The risk is that the technology becomes a gimmick, used because it is available rather than because it serves the game. The early applications of any new technology tend toward the gimmick, and conversational characters will likely go through the same phase. But beneath the novelty is a genuine change in what a game character can do, and the games that find the design language for it — rather than simply bolting conversation onto a scripted character — will be the ones that define what the technology ultimately means for the medium.