What is an AI-native game? A working definition

The definition

An AI-native game is one whose core experience depends on a model running while you play. Remove the model and the game doesn't just get worse; it stops being the same game.

That rules some things out. A game made with generative tools during development (concept art, placeholder voices, code completion) is AI-assisted. Players may or may not care about that, but it doesn't change what the game is. A game with a small "chat with this character" feature bolted onto a conventional design is AI-featured. A game where talking to characters is the play, where the world is generated as you explore it, or where agents run a society you live inside: that's AI-native.

Five design patterns

Most AI-native games we see combine one or two of these:

1. Characters you can talk to

NPCs that understand free-form text or speech and respond in character, often with memory of past conversations. Works best when talking is a meaningful verb: persuading, interrogating, negotiating, comforting. Fails when conversation has no consequences and the character is just a chatbot in costume.

2. Generative worlds inside authored constraints

Quests, levels, cases or regions created at runtime, but inside rules a designer wrote. The constraints are the craft. A murder mystery generator is only good if every generated case is solvable from the evidence, and that guarantee comes from hand-built logic, not the model.

3. Agent simulation

Many agents with goals acting on the world and each other, even when the player isn't watching: senates, villages, factions, markets. The appeal is emergent story; the risk is cost and incoherence. Most successful designs simulate in detail only what's near the player.

4. Adaptive directors

A system that watches how someone plays and adjusts pacing, difficulty, encounters or story beats. Classic games have had directors for years; models let them reason about style ("this party always scouts left") rather than just numbers. Players should be able to see and understand the adaptation, or it feels like cheating.

5. Natural language as input

Voice or text as the main way to act: orders in tactics games, spells described in words, building by description. The hard design problem is precision. Players need to know what's possible and get predictable results.

What separates good AI-native games

After reviewing hundreds of pitches and prototypes, the pattern in the ones that work is consistent:

• The model serves a mechanic, not the other way round. The best designs start with a verb ("lie to a suspect") and use a model because nothing else can do it.
• State lives outside the model. Inventory, quest flags, relationships and rules sit in deterministic code. The model reads and proposes changes; the game decides. That's how you keep things fair and debuggable.
• Heavy authorship. Detailed character bibles, tone guides, example dialogue and authored fallback lines. AI-native games often need more writing than conventional ones, just of a different kind.
• Consequences. What the player says changes something they can see later. Memory without consequence is a party trick.
• Graceful failure. When the model is slow, offline or produces something unusable, the game falls back to something authored instead of breaking.

A character that remembers you is a novelty. A character that remembers you and acts on it is a game.

Common pitfalls

• The blank-box problem. An empty text box intimidates players. Offer suggestions, show example lines, and make it clear what kinds of things work.
• Players breaking characters. Some players will try to make your characters say awful things or forget they're in a game. You need a tone layer, evals and red-teaming before launch, not after the first viral clip.
• Unbounded cost. Your most engaged players are your most expensive. Model costs per player-hour early; see our inference cost guide.
• Same-ness. Generated content can feel samey after hours of play. Vary it with authored seeds, strong character voices and structural variety the model can't flatten.
• Trust. Players are rightly suspicious of AI used to cut corners. Be clear about what's generated and why, and keep human craft visible. Our disclosure guide covers how.

How to prototype one in four weeks

The fastest way to learn whether an AI-native idea works is to build the smallest version that tests the core verb. Here's the four-week plan we suggest to studios before they pitch:

Week 1: one character, one goal

Build a single scene with a single character and one thing the player must get from them: a confession, a key, a vote. Hard-code everything around it. Use the biggest model you can afford; you're testing whether the interaction is fun, not whether it's cheap.

Week 2: consequences

Make what the player says change something: the character's trust, a door, a later scene. Move that state into code outside the model. This is usually where you find out whether the idea is a game or a conversation.

Week 3: break it

Give the build to people who weren't involved, and ask half of them to try to break the character. Log everything. You'll learn more about your guardrails, your blank-box problem and your fallback lines in one afternoon than in a month of design documents.

Week 4: measure

Count model calls and tokens per minute of play, and time every response. Build a first cost-per-player-hour estimate. Then write the one-paragraph pitch: what the player does, why it needs a model, what it costs. If that paragraph is convincing, you have something worth taking to a publisher.

Where it's going

Three shifts are making AI-native games much more viable than they were two years ago. Small models running on consumer devices have become good enough for focused characters, which makes per-player cost close to zero for many interactions. Latency from server models has dropped to conversational speed. And tooling for evaluating generated content (checking tone, consistency and rule-following at scale) has matured, which makes shipping to millions of players realistic rather than reckless.

We think the next five years produce the first AI-native genres, the way physics engines produced immersive sims and online play produced battle royales. That's the bet Ammighty is built on. If you're building one, we want to play it.

Frequently asked questions

Related

• AI-native games on the Ammighty slate
• LLM NPC inference costs
• Pitching an AI-native game