What AI-native games are, with examples you can play

Rick Vu, 9 October 2026

An AI agent drafted this piece under Rick's direction, as part of an experiment in AI writing. Every figure links to its source.

An AI-native game is a game that needs a generative AI model running during play for its core loop to work. Take the model out, or swap in something trivial like a fixed script, and the main way of playing collapses or turns into a different game. A game that only used AI to make its art or code before release does not count.

Where the definition comes from

The test used here is the one that AI Native Games: A Survey and Roadmap, a 2026 arXiv paper by Zhiyue Xu, Fandi Meng, Kaijie Xu, Clark Verbrugge, Simon Lucas, and Jian Zhao, adopts from Sun and colleagues’ 1001 Nights paper. The authors ask one question of each game. Is runtime generative AI “constitutive of the core loop”? In plain words, does output that a model produces while you play make up the thing you do over and over?

The test is counterfactual. You imagine the game without its model and check what is left. The authors screened candidates this way and kept 53 publicly available games and prototypes for analysis.

The survey also names two neighbors. An AI-augmented game runs a model during play, but only in a layer the game could lose, so the core loop survives if authored content or fixed rules replace the model. A boundary artifact puts the model at the center but has too little game structure to count as a game. Open-ended chatbot role-play sits there.

This piece is about games. The companion piece on what an AI-native game studio is extends the same test to the companies that make them. Rick Vu, who directed this draft, builds Red Dog Studio, a game studio building AI-native NPCs, and Scoat, an AI-native game maker.

Six kinds of AI-native game

The survey sorts games on two axes. One is the familiar game type, such as RPG, puzzle, or simulation. The other is the main job the model does during play. The second axis says more about what is new in each game. In plain terms, the six jobs are these:

• The model holds facts the player has to dig out. Investigation, interrogation, and lie detection games such as Vaudeville sit here.
• The model plays characters the player has to persuade, deceive, or bargain with. Suck Up! and 1001 Nights sit here.
• The model narrates. It takes open input and moves the plot, the relationships, and the story world, as in Whispers from the Star and Hidden Door.
• The model referees. It reads an action nobody planned for and decides what happens, as in Infinite Craft.
• The model runs a society of agents that the player watches or manages, as in Aivilization.
• The model makes images, text, audio, or objects that players then play with, as in the party game AI: Art Impostor.

The corpus is lopsided. The authors find it concentrated in language-forward designs. Referee games, agent simulations, generated construction, and companion play have fewer examples. Most AI-native games today are games you play by talking or typing.

What does not count

Most games that mention AI fail the test. They are a different and far more common thing.

A game with AI-generated art is not AI-native because of that art. The survey says a game “may contain AI-generated art, dialogue, or quests” and still not be AI-native if those outputs are static, optional, or replaceable by authored content. Textures a model drew before launch are fixed files. Delete the model after release and the game plays the same.

An optional AI feature is not enough either. If the player can turn the model off and play the same game, the model sits in the augmented layer. Procedural generation from hand-written rules, which has built roguelike levels for decades, is a separate tradition. The survey’s test asks about generative models.

Examples you can play

Each entry below is a released game or a public demo. For each one, ask what the model does while you play, and what is left if you take it away.

Whispers from the Star by Anuttacon came out on Steam on August 14, 2025, and has a demo. You help Stella, an astronaut stranded on an alien planet, by talking to her through voice, text, or video calls. The store page says “There’s no script” and that what you say shapes whether she survives. Swap the model for a dialogue tree and you get a different game, one where you pick lines someone wrote in advance. The survey files it under narration.

Suck Up! by Proxima is a comedy game on Steam, released October 1, 2025, in which you play a vampire. You speak out loud to AI neighbors and try to talk your way into their homes. Each one has its own personality and reacts to your tone and strategy. Without the model there is no one to persuade, and the whole loop is persuasion. In a developer post shared on the OpenAI community forum, one of its developers describes a “trust system” that balances difficulty without tying the model down.

Infinite Craft by Neal Agarwal is a free browser game on neal.fun that launched on January 31, 2024. You start with water, fire, wind, and earth, and you combine them in pairs. When nobody has tried a pair before, a language model invents the result. The game saves it, so every player gets the same answer for the same pair. Without the model, the game shrinks to the recipes already saved. The survey describes the model’s job as “judging and stabilizing concept combinations.”

AI Dungeon by Latitude calls itself “a free AI text adventure game where players create open-ended stories.” You type what your character does and the model writes what happens next. Remove the model and nothing remains. Yet the survey treats it as a boundary case, “closer to open-ended story continuation than to a strict game,” because it lacks stable goals and win or lose conditions. A model at the center is not enough. The game also needs structure around it.

inZOI is a life simulation that entered Early Access in March 2025 with Smart Zoi as an experimental setting. NVIDIA’s post on the inZOI and NARAKA launches says an on-device small language model of about half a billion parameters governs a Smart Zoi’s actions and thoughts. Players can also steer them with free-form text. The survey does not discuss inZOI. Applying its test, this article reads Smart Zoi as augmentation, because you can switch the setting off and still play the life sim.

Oasis by Decart goes the other way. A transformer generates every frame of an open world from your keyboard and mouse input, at 360p and 20 frames per second on NVIDIA H100 GPUs, according to Decart’s announcement. Decart says the model itself simulates the physics, the rules, and the graphics, so without it there is nothing on screen. The announcement linked a live web demo and released the weights of a smaller model to run locally. It also admits the model struggles to recall details from many frames back. The survey does not cover Oasis and lists world models only as a future direction. Whether a world with no goals and a short memory is a game is still an open question.

Why AI-native games are hard to design

The survey’s central design problem is how to turn open-ended language into stable gameplay. A model will accept anything a player types or says. A game needs a defined way to win, lose, and make progress. The authors list six things AI-native design still depends on: goals, rules, state, feedback, pacing, and player agency.

Goals come first. AI Dungeon shows what happens without them. The story continues, but nothing tells you whether you are winning. Suck Up! sets the goal outside the model. You get through the door or you do not.

Rules and state are where a model is weakest. A model can contradict itself from one turn to the next, and a game has to remember what is true. Infinite Craft settles this with a database. Once the model has answered a combination, the saved answer becomes the rule.

Feedback is how a player learns whether an action worked. Open conversation hides it. If a character warms to you, the game has to show that somehow, or you are guessing. A number the game tracks, such as a trust score, turns a conversation into progress the code can check.

Pacing runs into speed. The survey notes that “current model inference is still often too slow and variable to support tightly timed feedback loops.” Typed conversation also creates what the authors call a “fragile turn rhythm.” That is probably one reason the corpus leans toward talking games and away from action games. One survey example, Historical Simulator: Chongzhen, works around it with an edict and memorial structure. You issue orders, and the game updates its state at intervals.

Agency comes last. The survey points out that when a model forgets earlier history, the player has to supply it again, which “undermines agency.” A player who explained something to a character an hour ago expects the character to know it. In an AI-native game, memory is a design decision as much as an engineering one.

Related reading

• How AI NPCs work, for what runs inside one of the characters in a conversation game.
• World models and games, for more on Oasis and the generated-world approach.
• AI-native startups in the gaming market in 2026, for the companies building these games.