What an AI-native game studio is and how to test for it

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 studio is a studio whose games depend on generative AI that runs while people play. Remove the model and the core of the game stops working or turns into a different game. A studio that uses AI to write code or draft concept art is a studio that uses AI tools, and by that measure much of the industry already qualifies.

Nobody has defined the studio yet

Funding announcements and vendor posts often use “AI-native” as a label for a small team with a new pitch, and they rarely say how to check whether a given studio fits.

The closest thing to an authority defines games, not studios. AI Native Games: A Survey and Roadmap, a 2026 arXiv paper, calls a game AI-native when runtime generative AI is “constitutive of the core loop.” Its test is counterfactual. If you remove the AI or replace it with something trivial, the central form of play collapses or changes fundamentally. The authors apply that test to 53 public games and prototypes, and they find the corpus leans toward designs built on language.

That test is the right place to start. It asks about the player’s experience instead of the team’s workflow, and it gives a yes or no answer for a single game. A studio needs one more step.

Start with the game

Take two games.

In the first, you question a suspect by typing whatever you want. The suspect answers in character, lies when a lie protects them, and slips when you press on the right detail. Replace the model with a dialogue tree and you have a different game. The player now picks from a menu of questions someone wrote in advance. The skill the game asked for, which was finding the right question in your own words, is gone. This game passes.

In the second, you clear rooms in a shooter while enemies shout lines that a model wrote during production. Replace those lines with lines a writer wrote and recorded, and nothing about play changes. This game fails the test, however much AI went into making it.

The amount of AI the team used does not decide the answer. What decides it is whether a model still runs when the player plays.

Extend the test to the studio

A studio uses generative AI in many places, and only some of them matter for this label. Sort every use into one of three layers.

Generated at runtime

The model runs on a server or on the player’s device during play. Its output reaches the player without a person reviewing it first. A character that writes its reply on the spot is runtime generation. So is a quest that a model assembles from what the player has done. Only this layer can make a game AI-native, because only this layer can fail the counterfactual.

Generated at build time

The model runs before release. A person reviews the output, and the shipped game contains fixed files. Textures, voice lines, level layouts, and code written with an agent all belong here. If you deleted the model after launch, the game on the store would not change.

AI-assisted

The model helps a person think or work, and nothing it produces ships directly. Research, brainstorming, and code review fall here. The GDC 2026 State of the Game Industry report found that 36% of game industry professionals use generative AI tools in their job, and that research or brainstorming is the most common use.

What makes a studio AI-native

With the three layers in place, a definition for the studio follows. An AI-native game studio is one whose main games pass the runtime test, and which organizes its work around that runtime layer. Build-time and assisted uses can be as heavy or as light as the team likes. They do not move a studio into the category or out of it.

The second half of that definition carries weight. A studio that ships one experiment with a model in the loop has made an AI-native game. It has not yet become an AI-native studio. The label fits when the runtime model changes how the studio works every day. Its designers write rules for what a character may and may not say, instead of writing every line. Its testers check output they have never seen, because the model produces new output for each player. Its budget counts inference cost for every hour someone plays, a cost that a studio shipping fixed files never pays.

Using AI tools does not make a studio AI-native

The layers also explain why the word gets stretched. Build-time and assisted AI are now common. Dataconomy’s report on Steam Next Fest demos counted 1,704 of 8,700 titles in the June 2026 Next Fest, or 19.5 percent, with a generative AI disclosure. That figure does not tell you which layer the AI sits in. A disclosure on a store page covers a generated texture and a live model alike.

The GDC 2026 State of the Game Industry report also found that 52% of game industry professionals think generative AI is having a negative impact on the industry. The definition here keeps that debate apart from the label. A studio can hire illustrators, draw every frame by hand, and still be AI-native if a model runs its characters during play. Another studio can generate every texture with a model and still fail the test. Whether to use generative art is a separate decision on a separate layer.

Edge cases

Chat NPCs

A chat window on a character does not, by itself, make a game AI-native. Apply the counterfactual. If the player can ignore the chat and finish the game the same way, the model is decoration. The conversation has to change something the game tracks, such as what a character knows or which path opens. Then removing the model removes a way to play.

Writers stay in the loop too. Ubisoft’s NEO NPC prototype, built with Inworld and NVIDIA and shown at GDC 2024, still had writers author each character’s persona, according to VGC’s report on Ubisoft’s AI-powered characters. The runtime layer changes what writers write. It does not replace them.

Procedural generation

Games have built levels from rules and random seeds for decades. That is generation at runtime, but it is not generative AI in the survey’s sense. A studio built on hand-written generators is a procedural studio. The line between the two is not sharp. This article draws it at learned models, because that is where output stops being something the designer can enumerate.

Tools for making games

Some AI-native companies make tools instead of games. The test still works with one change. Remove the model and ask whether the product still does its main job. If it cannot, the tool is AI-native. The games people make with it are a separate question, and each one gets its own test.

Where Rick’s companies sit

Rick Vu, who directed this article, runs two companies this definition covers, so read this section with that interest in mind. Red Dog Studio builds AI-native NPCs, which places its work in the runtime layer. Scoat is an AI-native game maker, which puts it in the tool case above. This article does not score either one against the test, because the details the test needs are not public.

Apply the test yourself

You can run this on any studio with its public trailers and store pages.

1. Write the game’s core loop in one sentence. “Question suspects until you can name the killer” is one.
2. Mark each part of that loop that a model produces while the player plays.
3. Replace each marked part with a fixed script, a lookup table, or nothing. Ask whether the sentence from step 1 still describes the game.
4. If the sentence no longer fits, the game is AI-native. Every other use of a model in that game is build-time or assisted work.
5. For a studio, run steps 1 to 4 on the games it has shipped and the one it is building now. If the runtime layer sits at the center of most of them, the studio is AI-native.

A studio that fails step 5 can still be good at using AI. It is a studio with AI tools, which is a different and much more common thing.

Related reading

• How AI NPCs work, for what the runtime layer looks like inside a single character.
• AI game maker vs game engine, for the tool case in more detail.
• Game design lifecycle vs software lifecycle in the age of AI coding, Rick’s essay on how AI coding changes the order of work.