Survey Defines AI-Native Games, Outlines Future Development Roadmap.

The proliferation of generative AI has enabled games to dynamically create elements like dialogue, quests, characters, and worlds. However, simply using generation doesn't make a game "AI-native." This research proposes a rigorous definition: an AI-native game is one where the core gameplay would fundamentally collapse or change if the generative AI component were removed or trivially replaced. This criterion helps differentiate truly AI-native experiences from those merely augmented by AI. The paper presents a comprehensive survey of 53 publicly available AI-native games and prototypes, introducing a dual-axis G/N taxonomy. The G-axis categorizes player-facing game types, while the N-axis focuses on the dominant AI mechanic essential to play. The analysis reveals a concentration in language-forward designs, such as narrative adventure and epistemic interaction, with other categories like multi-agent simulation and generative construction being less explored. A key challenge identified is organizing the inherent "semantic openness" of AI outputs into stable, engaging gameplay. Successful AI-native design relies on maintaining mechanical invariants like goals, rules, and player agency, which make the open-ended AI outputs interpretable and consequential. The paper concludes with a roadmap addressing critical areas for future development, including controllable generation, AI-as-mechanic design, multimodal systems, inference economics, evaluation, safety, and regulation.