AI Native Games: Survey, G/N taxonomy and research roadmap

Zhiyue Xu and five co-authors submitted a paper on 1 July 2026 that defines "AI-native" games and analyzes 53 publicly available AI-native games and prototypes. The authors introduce a dual-axis G/N taxonomy, identify where current work concentrates, and outline a roadmap covering controllable generation, multimodal systems, evaluation, safety, and regulation.

What is an AI-native game?

The paper answers this by offering a counterfactual criterion: an AI-native game is one where runtime generative AI is constitutive of the core loop, so "if the AI component were removed or trivially replaced, the central form of play would collapse or become fundamentally different." This separates AI-native games from AI-augmented titles, tavern-style role-play, procedural content generation used only in production, and chatbots.

The authors frame the definition to focus on indispensability at runtime, not merely the presence of generated assets. That lens shifts evaluation from whether a game uses generative models to whether the generative component is essential to what players do and experience.

How did the authors classify existing AI-native games?

They screened and analyzed 53 publicly available AI-native games and prototypes and placed them in a dual-axis taxonomy: the G-axis captures player-facing game type and the N-axis captures the dominant AI mechanic that makes generative AI indispensable to play. The corpus is concentrated around language-forward designs, with particular emphasis on narrative adventure, epistemic interaction, and generative narrative.

By contrast, the paper finds categories that remain less represented: semantic adjudication, multi-agent simulation, generative construction, and relationship or companion play. The taxonomy highlights both what designers are building now and gaps where further work could expand the space of playable AI-native designs.

What core design problems did the survey identify?

The central design problem the authors identify is how to organize semantic openness into stable gameplay. They argue that AI-native design depends on mechanical invariants that make open-ended AI outputs interpretable and consequential. Those invariants include goals, rules, state, feedback, pacing, and player agency.

The paper treats these invariants as necessary scaffolding: without them, open generative outputs risk producing noisy, unplayable, or ambiguous interactions. The authors use this diagnosis to motivate research directions that focus on controllability and on treating AI as a game mechanic rather than an add-on.

Why it matters

The definition and corpus shift the conversation from using generative models as production tools to designing games where generative AI is the gameplay engine. By isolating a counterfactual criterion and cataloguing 53 public examples, the paper gives designers a practical test for whether an AI feature is core or peripheral. That clarity affects design choices, evaluation metrics, and regulatory thinking for interactive systems that rely on runtime generation.

This focus on mechanical invariants also reframes technical work: progress in controllable generation, multimodal and multi-agent systems, and inference economics will directly enable more reliable and interpretable AI-native experiences.

What to watch

Track research and tools that improve controllable generation and explicit mechanisms for AI-as-mechanic design, plus publications or prototypes that target the less-represented categories the authors highlight. Also watch for new evaluation methods and safety or regulation discussions tied to runtime generative systems, which the paper lists as part of its roadmap.

Authors: Zhiyue Xu, Fandi Meng, Kaijie Xu, Clark Verbrugge, Simon Lucas, Jian Zhao. Submission date: 1 July 2026. Analysis sample: 53 publicly available AI-native games and prototypes.