Theory — How AI Creates Games (and Why It Fails)

You've played. You've reflected. Now you understand *why* AI is brilliant as a game master — and at the same time a poor game designer.

The Brain of an AI Game Master — Tokens and Context

Remember: AI works by predicting the next word.

That's also how AI runs a game.

You say: "I open the door."

The AI has read this sequence: [Protagonist] [Action: open] [Object: door]

The AI asks: What comes next? Which words follow "I open the door"?

The AI has read millions of stories where someone opens a door. It sees:

• 20% chance: "You see a dark corridor"
• 15% chance: "An NPC waits behind it"
• 10% chance: "It was a trap"
• 10% chance: "There's nothing behind it"
• etc.

The AI picks one of these options (weighted by probability) and writes the next words.

This is not thinking. This is statistical prediction.

But here's what's interesting: This simple principle makes it a brilliant storyteller. It knows which scenes fit together. It knows which dialogue sounds natural. It knows what feels surprising.

That's its superpower: Narrative Plausibility.

The Context Window — Understanding the Memory Problem

But there's a catch: The context window.

A context window is like a notebook where the AI writes down everything that's happened so far. In ChatGPT that's about 4,000 words. In Claude about 100,000. In future models maybe more.

That sounds like a lot. But in a game it's not far enough.

Imagine: You play 20 turns. Each turn is about 100 words. That's 2,000 words. You have 2,000 words of context left.

Now the AI had to remember thousands of details:

• The NPC's name
• The color of his shirt
• That you have a coin
• That the forest is to your right
• That the well was poisoned
• That you made a pact with the merchant
• That the door is made of iron
• etc.

By turn 15? Gone.

This is not stupidity. This is architecture.

The Three Task Types in Games — The Pattern

Here it gets practical. Think of the types from L03 Text, but applied to games:

Type 1: The Multiplier

You have the game concept, AI does the details.

You say: "I want a wizard duel. Here are the rules: Each wizard has 100 mana. Fire spells cost 20 mana and do 30 damage. Water spells cost 15 mana and heal 20 HP. My character is a fire wizard with 50 HP. Let's fight."

Now the AI can follow the rules. It will remember the numbers (at least for a while). It will move the wizards. It will describe the effects.

AI is great here because the rules are external. It doesn't have to invent them — it just has to execute them.

That's the multiplier: Fast, consistent execution of a system someone else designed.

Type 2: The Enabler

You don't know game rules, AI invents them for you.

You say: "I want a simple adventure. A character explores a ruin. Whenever I enter a location, you describe what's there. I choose what I do."

Here everything is up to the AI. It has to invent the world. It has to invent the logic. It has to react to your moves.

And it can do that! That's exactly what AI does well — fast, responsive improvisation.

That's the enabler: Someone with no experience plays a game the AI spontaneously invents.

Type 3: The Limits

This doesn't work with AI.

You say: "I want a tactical game. There's a 10×10 map. I'm at square (3,5). The enemy is at square (8,8). We move alternately. The enemy uses optimal strategy to catch me."

Here the AI fails.

Why? Because a tactical game requires spatial reasoning. The AI has to visualize the map, calculate distances, compute the best strategy.

But AI doesn't think spatially. It thinks narratively. It doesn't say "the optimal move is north-west". It says "the enemy moves toward you".

Also: If the AI thinks the strategy, it can't play a "real" strategy. It's theatre. The enemy will win or lose depending on what the AI thinks is exciting — not because it's tactically better.

That's the limit: Complex systems, spatial thinking, real balance.

Why AI Is Great at Open Worlds but Not at Rule Systems

That's the core difference:

An open world (like a fantasy ruin) works with narrative plausibility. The AI only has to ask: "What's probably what I should describe here?" And it answers right.

A rule system (like chess or a combat system with numbers) needs exact logic. The AI has to keep numbers in mind, apply formulas, think strategically.

AI can do the first. It can simulate the second, but not really.

A Concrete Example — Understanding This

Scenario 1 — The AI nails it:

You enter a cursed castle.

You: "I look around."

AI: "The walls are overgrown with ivy. An old chandelier hangs from the ceiling. On the floor you see scratches, as if something heavy lay there."

This works beautifully. The AI imagines an abandoned castle and describes details. Pure narrative.

Scenario 2 — The AI fails:

You play a number puzzle.

You: "I see 3 switches. The first is red, the second blue, the third green. If I push the red one, the door opens 1 meter. The blue one pulls it 0.5 meters closed. The green one pushes it 2 meters open. What order solves the puzzle?"

Now the AI has to calculate. It has to think logically. This is not narration, this is mathematics.

And that's where it gets difficult. The AI will guess an order that sounds probable. But it won't be guaranteed correct.

Consequences — How to Use AI Right

This means practically:

Use AI for:

• Open, narrative experiences (fantasy adventures, mysteries, dramas)
• Fast, detailed descriptions (scenes, NPCs, dialogue)
• Improvisation on player input (react spontaneously to unexpected moves)
• Co-narration (you design the structure, AI tells the details)

Avoid:

• Strict rule systems (AI will make mistakes)
• Spatial reasoning (AI doesn't think cartographically)
• Long-term memory (after 50+ turns it gets chaotic)
• Real strategic balance (AI can't play a fair duel)

The Big Thought — What AI Really Is in Games

Here's the central insight:

AI is an interactive novel generator, not a game engine.

A novel needs no rules. A novel is: Description → Action → Reaction → new description.

A game needs a system. The system makes sure it's fair, consistent, challenging.

AI can deliver the first. A human has to build the second.

That's where you realize the boundary isn't in creativity — it's in architecture.

A Thought: The Perfect Game with AI

What's the best way to play a game with AI?

It's not: AI tells the whole game.

It's: You are the rule engine. AI is the narrator.

You make the rules. You track the numbers. You ensure balance. You decide if the quest is hard enough.

AI tells the story. It paints pictures. It fills roles. It reacts to your moves.

With this division it becomes brilliant. And that's exactly what we'll practice in L04.

Summarized

• AI works by predicting the next word (token).
• The context window limits memory to about 4,000–100,000 words.
• Multiplier: You design the rules, AI executes fast.
• Enabler: AI spontaneously invents a game that works.
• Limits: Rule systems, spatial thinking, balance.
• AI is an interactive novel generator, not a game engine.
• The perfect game: You're the engine, AI is the narrator.