AI-native Games

• AI-native games are defined by their reliance on generative AI as the essential mechanism, where removal of the AI disrupts the core gameplay loop.
• They employ dynamic content generation, adaptive NPC behavior, and personalized play loops to create innovative, research-driven gaming experiences.
• Benchmark frameworks and competition platforms demonstrate technical intricacies and educational applications, advancing both game design and AI research.

AI-native games are games in which AI/ML is not a peripheral tool but an essential substrate for gameplay, content creation, NPC behavior, live operations, and player personalization; in the generative-AI literature, an AI-native game is one in which runtime generative AI is a constitutive core mechanism of play, such that if the model were removed or trivially replaced, the core loop would collapse or become fundamentally different in kind (Karpouzis et al., 2021, Xu et al., 1 Jul 2026). The term also appears in a benchmark-oriented sense: a game or competition framework may be deliberately designed as an AI-native game and competition framework when simple deterministic rules, clean standardized interfaces for AI agents, scalable parameters, and open tooling make it simultaneously a research testbed and an educational vehicle (Fachada, 2020).

1. Definition, boundary conditions, and the counterfactual test

The contemporary definition is anchored by a counterfactual criterion. Runtime generative AI must participate during play; model outputs must directly affect action interpretation, produced content or consequences, NPC behavior, world state, or outcome judgment during play; and the same core experience must not be preservable by swapping in finite authored assets or a deterministic mechanism without changing the central form of play. In formal restatement, let a game be $G=(S,A,T,R,L)$, with state $S$, player actions $A$, transition or update $T$, reward or goal $R$, and core loop $L$. An artifact is AI-native if and only if the runtime model $M$ influences $T$ or $L$, and if replacing $M$ with a finite, deterministic substitute $S$0 makes $S$1 undefined or yields a qualitatively divergent play distribution, $S$2 (Xu et al., 1 Jul 2026).

This criterion separates AI-native games from AI-augmented games, boundary artifacts, chatbots, tavern-style role-play, procedural content generation, and AI-assisted production. It also clarifies an HCI-oriented formulation in which the core loop requires direct, repeated interaction with AI capabilities—training and steering models, co-creating or conversing with them, or competing against them—so that the experience would be materially impossible or categorically different without the AI. In that sense, “AI as play” expands human–AI interaction beyond productivity tools and makes the model itself a first-class play object (Zhu et al., 2021).

Category	Generative AI the core mechanism?	Strict game with ludic structure?
AI-native	✓	✓
AI-augmented	✗	✓
AI-boundary	✓	✗
Out of scope	✗	✗

A recurrent misconception is that generation alone makes a game AI-native. The survey literature states the opposite: generation alone does not make a game AI-native, nor does it guarantee playability. The decisive issue is whether the runtime model performs indispensable mechanical work inside the core loop (Xu et al., 1 Jul 2026).

2. Intellectual lineage: from AGI testbeds to open-ended game distributions

A major antecedent is the use of games as environments for measuring intelligence. Schaul, Togelius, and Schmidhuber argued that games are suitable for measuring general intelligence because they combine simplicity and controllable complexity, natural structured rewards, diversity of cognitive demands, procedural generation and reproducibility, and speed and availability. Their starting point is the Legg–Hutter definition,

$S$3

followed by a finite-time, practical extension in which environment complexity is approximated by

$S$4

and performance is estimated by Monte Carlo sampling over a restricted game class (Schaul et al., 2011).

This line of thought later shifted from classic board games to general video game playing. Video games were described as “simpler, cheaper and faster than robots,” able to run in controllable environments, to be sped up “to many times the original speed,” and to support “experiments… many thousands of times in quick succession.” The General Video Game Playing Competition and the Video Game Description Language operationalized the demand for many, diverse, unseen games, while general video game generation pursued an “endless supply of new games” to reduce per-title tuning and game specificity (Togelius, 2016).

A recent large-scale continuation is AI GameStore, which reframes evaluation around the “Multiverse of Human Games”: the open-ended space of all conceivable human games plus its cultural distribution. AI GameStore uses LLMs with humans-in-the-loop to synthesize representative human games from popular platforms, standardizes them in p5.js, annotates them across Visual Processing, Spatial-temporal Coordination, Memory, Planning, World Model Learning, Physical Reasoning, and Social Reasoning, and evaluates humans and models through a common harness. In a proof of concept, 100 such games were generated and seven frontier vision-LLMs were evaluated on short episodes of play; the best models achieved less than 10% of the human average score on the majority of the games and especially struggled with world-model learning, memory and planning (Ying et al., 19 Feb 2026).

3. Taxonomies, mechanics, and the problem of semantic openness

The most extensive recent corpus study screened 93 candidates and retained 53 publicly available AI-native games and prototypes. Its dual-axis taxonomy distinguishes a G-axis of player-facing game type from an N-axis of the dominant AI mechanic indispensable to play. On the G-axis, the corpus contains G1 Narrative Adventure (24; 45.3%), G2 RPG (8; 15.1%), G3 Puzzle (7; 13.2%), G4 Strategy/Management (1; 1.9%), G5 Simulation (4; 7.5%), G6 Sandbox/Creation (2; 3.8%), G7 Social Deduction/Party (3; 5.7%), G8 Relationship/Companion (2; 3.8%), and G9 Hybrid/Experimental (2; 3.8%). On the N-axis, it contains N1 Epistemic Interaction (17; 32.1%), N2 Social Influence (11; 20.8%), N3 Generative Narrative / AI GM (14; 26.4%), N4 Semantic Adjudication (6; 11.3%), N5 Multi-Agent Simulation (4; 7.5%), and N6 Generative Construction (1; 1.9%). Dense regions are G1×N1, G1×N3, and G2×N3; underexplored regions include G8 Relationship/Companion and the N6 Generative Construction column (Xu et al., 1 Jul 2026).

The broader AI-in-games taxonomy helps explain why these concentrations occur. AI can be integral through co-creative content systems that generate levels, quests, music, and dialogue on the fly based on player behavior and affect; adaptive NPCs whose plans, tactics, and social responses change with player actions and cultural contexts; runtime personalization engines that monitor mastery and preferences to select activities and recalibrate difficulty; and live-ops analytics that process player logs and communication to detect attrition risk, toxicity, or sentiment and trigger in-game interventions. The same chapter organizes the field into AI for gameplay, AI for content, AI for analytics, and games for AI as data sources and benchmarks (Karpouzis et al., 2021).

The central design problem is organizing semantic openness into stable gameplay. The survey identifies “mechanical invariants” that make open-ended AI outputs interpretable and consequential: goals, rules, state, feedback, pacing, and player agency. This makes semantic adjudication especially significant. In N4 designs, the model acts as a rule interpreter or semantic physics engine, but proposal–verification loops, structured outputs, function calls to engine actions, validators, and caching are required if free-form language is to remain legible, fair, and mechanically stable (Xu et al., 1 Jul 2026).

4. Benchmarks, competitions, and agent-facing exemplars

One strand of AI-native design is explicitly agent-facing. ColorShapeLinks was deliberately designed as an AI-native game and competition framework. Its core game is an arbitrarily sized variant of Simplexity that couples simple, fully deterministic rules with deep strategic structure. The game is deterministic, two-player, zero-sum, turn-based, and of perfect information; it is parameterized by board rows $S$5, columns $S$6, win sequence length $S$7, and per-player inventories $S$8 and $S$9, with defaults $A$0, $A$1, $A$2, $A$3, $A$4. Its turn bound is

$A$5

its state-space upper bound is $A$6, its initial branching factor is $A$7, and an upper bound on the game tree size is $A$8. The framework is implemented in C# with a Unity application and a .NET console application, exposes the AbstractThinker agent interface, ships with Sequential, Random, and Minimax baselines, and supported both internal class competitions and the IEEE Conference on Games. In the CoG 2020 competition, Thunder won the Base Track, SureAI won the Unknown Track, and SimpAI was runner-up in the Unknown Track (Fachada, 2020).

The Text-Based Adventure AI Competition is an AI-native benchmark around the epistemic challenges of acting in non-operationalized environments. Instead of enumerated action sets, agents receive text-only observations and return unrestricted text strings; a convenient formalization is a POMDP with textual observations $A$9, textual commands $T$0, latent game states, and a policy over dialogue history. Across the standardized corpus of 20 games, NAIL achieved mean completion $T$1 and success $T$2 at 1,000 steps per game, while persistent failures on score parsing, meta-dialogues, sparse rewards, and affordance discovery showed that action generation, world modeling, and exploration remain tightly coupled challenges (Atkinson et al., 2018).

A still earlier instance is AI WAR, described as a 3D game and an engine to design and develop war machines using artificial intelligence. Players author autonomous Cybugs in the CAICL language, and the human player acts as a coach: strategy is expressed through condition-action rules over scans, collisions, fuel, and damage, then observed in autonomous combat. The sense–decide–act loop—scanning, movement coordination, dynamic resource allocation, opponent modeling, and planning or problem solving—makes gameplay emerge from programmed behavior rather than direct control (Ahmed, 2010).

These benchmark and competition forms sit within a wider design space of player–AI interaction. A survey of 38 neural network games identified four dominant metaphors: NN as Apprentice (34%), NN as Competitor (26%), NN as Teammate (21%), and NN as Designer (19%). The survey’s implication is that AI-native games need not be only competitions or sandboxes for autonomous agents; they also include co-creative loops, adaptive adversaries, and training-as-play systems in which the player learns the model by manipulating it (Zhu et al., 2021).

5. Runtime architectures, generative pipelines, and learned engines

Generative AI-native games often place a runtime model directly in the content loop. “Malinowski’s Lens” defines AI-native play as one in which AI models are at the center of gameplay and content creation. The system transforms Argonauts of the Western Pacific into an interactive learning experience through Retrieval-Augmented Generation and DALL·E 3 text-to-image generation. The corpus is embedded with OpenAI text-embedding-3-small, stored in ChromaDB, orchestrated through LangChain, generated with GPT-4o, served by FastAPI, and rendered in a React single-page app built with Vite. Every turn’s narrative and visuals are generated in real time; AI is the primary driver of interaction, narrative, and player engagement. Reported turn latency is approximately 40–50 seconds, and total reported cost across development and playtest phases is €33 (Hoffmann et al., 10 Nov 2025).

At a different layer, PlayGen treats the engine itself as a learned model. Its runtime simulator approximates $T$3 through a VAE, an autoregressive DiT-based latent diffusion model, and an RNN-like hidden state that summarizes temporal context. The system was validated on Super Mario Bros and Doom, runs at 20 FPS on an NVIDIA RTX 2060 GPU at 128×128, and sustains these results even after over 1000 frames of gameplay. Its evaluation framework introduces a Valid Action Model and two action-aware mechanics metrics:

$T$4

and

$T$5

The paper states that ActAcc $T$6 and ProbDiff $T$7 correlate with human judgments of correct mechanics (Yang et al., 2024).

A third architectural direction concerns agent generalization across games and rule sets. “AlphaViT,” “AlphaViD,” and “AlphaVDA” replace AlphaZero’s fixed-size trunk with Vision Transformer tokenization and, in the decoder variants, adaptive policy heads whose action-space size can change at runtime. A single network with shared weights can play multiple board games of various sizes; AlphaViT employs only a transformer encoder, whereas AlphaViD and AlphaVDA incorporate both transformer encoders and decoders. Experimental results show that these agents consistently outperform traditional algorithms such as Minimax and Monte Carlo Tree Search and approach the performance of AlphaZero, while simultaneous training on multiple games yields performance comparable to, or even surpassing, single-game training (Fujita, 2024).

These architectures indicate that AI-native play can locate “the AI” at different levels: in narrative grounding and rendering, in the simulator or engine, in the policy that generalizes across variable board sizes, or in the adjudication layer that maps open input into valid state transitions. This suggests a family resemblance rather than a single implementation pattern.

6. Evaluation, educational use, tensions, and future directions

AI-native games are increasingly used as educational systems as well as research artifacts. “Malinowski’s Lens” reports two validation studies: Study 1 with 10 non-specialists showed average quiz score 7.5/10 and SUS 83/100, while Study 2 with 4 expert anthropologists confirmed pedagogical value and included a senior researcher who discovered “new aspects” of Malinowski’s work through gameplay. The game’s dual-phase structure—fieldwork exploration followed by an academic defense with a tailored 10-question quiz—illustrates one way to convert retrieval-grounded generation into assessment (Hoffmann et al., 10 Nov 2025).

ColorShapeLinks was integrated into two AI-for-Games course offerings at Lusófona University. In semester 1, passing required beating the sequential and random agents; in semester 2, the console front end and the baseline minimax agent were added, and passing required beating minimax. Reported outcomes were Semester 1 $T$8: mean $T$9, median $R$0, pass rate $R$1; Semester 2 $R$2: mean $R$3, median $R$4, pass rate $R$5. Several teams adopted transposition tables, parallelization, and heuristic tuning, and students published follow-up work such as SimpAI (Fachada, 2020).

A more explicitly pedagogical AI-literacy line proposes “Learn Like an LLM” and “Tag-Team Text Generation.” These games operationalize maximum-likelihood training on a dataset and autoregressive next-token sampling with randomness through simple mechanics, immediate feedback, and structured debriefs. The paper is explicit that the designs are still in early stages and presents no formal evaluation; the importance of the contribution lies in embedding negative log-likelihood, probability distributions, temperature, and sampling into rules rather than merely theming a game around AI (Chen et al., 30 Mar 2026).

Evaluation remains a central difficulty. The recent survey argues for playability-first evaluation rather than model benchmarks alone and identifies rule consistency and fairness, narrative coherence and memory stability, player agency and feedback quality, latency and responsiveness, safety and moderation success, and replayability as core criteria. It also argues that the main design challenge is to convert open-ended, stochastic AI outputs into play that remains interpretable, fair, and consequential (Xu et al., 1 Jul 2026).

The literature is equally explicit about limitations. Runtime generation alone does not guarantee playability; latency harms turn rhythm in dialogue-heavy loops; inference economics shape feasible designs; and safety, privacy, bias, hallucinations, prompt injection, and manipulation risks are heightened by runtime generation. In educational and live-service contexts, the relevant concerns include privacy-preserving capture, opt-in and anonymization, mitigation of toxic behavior, subgroup performance and error rates, and transparency of adaptivity and data use (Karpouzis et al., 2021). More specific limitations include the narrow focus and possible shelf-life compression of a single benchmark game such as ColorShapeLinks once near-perfect agents or analytical solutions emerge, the anthropologist-centered viewpoint and exclusion of indigenous voices in “Malinowski’s Lens,” and the difficulty of sustaining reproducible evaluation when hosted models, prompts, and APIs change over time (Fachada, 2020, Hoffmann et al., 10 Nov 2025).

The near-term roadmap is correspondingly technical rather than rhetorical: controllable generation via proposal-and-verification pipelines; AI-as-mechanic design that binds open input to explicit state changes; multimodal and multi-agent systems; memory inspection, repair, and caching; local or quantized models, batching, and speculative decoding to manage cost and latency; long-session, adversarial, and memory-stress evaluation; and model-agnostic preservation layers that record prompts, schemas, model versions, logs, and safety incidents (Xu et al., 1 Jul 2026). A plausible implication is that the field’s maturation will depend less on adding more generation and more on stabilizing the interfaces between semantic openness and mechanical invariants.