What makes a game AI-native?

If you replaced a game’s language model with written replies, could players still play it the same way? That question gives you somewhere to start when deciding whether to call a game AI-native.

For Xu and colleagues, an AI-native game needs generative AI to run during play and shape the core loop. Replacing it with written content or fixed rules would change how the game plays. That is the survey’s working definition, not an agreed industry standard.^[1]

Take a hypothetical detective game. A suspect answers questions in different words each time, but always gives up the same clue. You could write those answers in advance. The player would still get the clue and move on with the case.

The question gets harder if the player can invent an argument that changes what the suspect reveals. Replacing the model now means deciding which arguments count and what each one does. Some of the player’s options may disappear when you write those rules.

You could check this in a single scene. Keep a record of what the player asks and which clues become available, then work through the scene with written replies. A different turn of phrase is easy to spot. To understand the effect on play, trace whether the player can still get the same information and take the same next step.

There is also a boundary to the term “game” here. The survey looks for goals, rules, and changes to game state that have consequences. A conversation that depends on a model still needs to meet that boundary. The classification says nothing by itself about whether the game is enjoyable.^[1]