AI Can't Read Fun

My Sleep book is out! Get your copy from Sleepagain.co.

And now for this week’s piece.

In January 2026, Steam updated its AI disclosure form. Developers now answer three questions instead of one: does your game use generative AI at all, does it pre-generate content that ships with the game, and does it live-generate content during gameplay. Coding helpers are exempt. The focus is on what the player actually sees.

One in five new Steam releases now carries an AI disclosure. That number will keep rising.

I run an AI game development course, and one of the lessons walks through level design specifically. The Steam policy maps almost exactly to the three modes I teach: pre-generated levels, real-time levels, and player-generated levels. Each mode has a different relationship with AI, and each has a different answer to the same underlying question.

That question is: can AI design fun levels?

Raph Koster gave the clearest answer I’ve read in A Theory of Fun for Game Design. His argument is that fun is the brain’s reward for learning. Humans are pattern-recognition machines. Games present patterns to master, and when your brain is learning, it releases a reward signal. That signal is what we call “fun.”

AI can handle patterns. It can build rule systems, calculate constraints, and validate that a level is solvable. What it cannot do is feel whether the learning curve is right. It cannot play your level and know it’s boring at minute eight. That part still requires a human.

This split, math on one side and creativity on the other, is why the three modes of level design behave so differently in practice.

These are the levels the designers build during development and ship inside the game. There is a linear path, difficulty spikes, and new mechanics that get introduced as the player levels up in skill and gear. AI is extremely useful here. You can generate dozens of level variants that satisfy a constraint set, validate them for solvability, and let a human designer pick the ones that feel right.

The key insight is that AI does the math. The human still does the judging. If you skip the human pass, you ship boring levels that technically work.

These get generated while the player plays. The game reads what the player is doing and builds the next room, the next wave, the next puzzle on the fly. In theory this is personalization. In practice it is hard, because AI does not know what “fun” looks like for this specific player in this specific moment.

Most real-time level generation that works well today is not pure AI. It is a rules engine with AI assistance. The game has a library of hand-tuned building blocks and uses AI to arrange them. Same pattern again: AI does the combinatorics, humans designed the blocks.

This is the mode I find most interesting. Players have always wanted to build their own levels. The problem is that level design tools are cumbersome. You need to learn the editor, understand the constraints, and spend hours iterating.

With AI, players can describe what they want in a prompt. “A desert map with three choke points and a hidden shortcut through the canyon.” The AI builds a first pass. The player tweaks it. The cycle that used to take two hours takes ten minutes.

This lowers the barrier for creativity without removing the creativity. The player is still the one reading whether the map is fun, because they play it.

The clearest example I have of the math-versus-creativity split is outside of games. It is a crossword puzzle project I’ve been running since summer 2025.

I wanted a 5x7 grid with clues, valid words, and proper crossings. I tested the same prompt against ChatGPT, Gemini, and Claude. All three struggled. In Finnish, they invented words that do not exist. In English, they produced grids with mismatched letters at the crossings. Single prompts were never going to solve it.

In the fall, I built three separate projects pointed at how professional crossword constructors actually work, step by step. None of them produced a valid puzzle.

The thing that eventually worked is Claude Code with a set of guard rail scripts. The scripts do the deterministic work: check the grid, flag duplicate clues, enforce word-length constraints, reject words I’ve already used in earlier puzzles. Claude does the search and the filling. When Claude tries to take a shortcut, the script catches it and tells Claude exactly what to fix. Claude Code became the software operator, not just a code generator.

That framework has now built 140 puzzles for a couple’s app I’m working on. Each puzzle takes 10 to 15 minutes. I usually kick off a run and go do something else. The grids are valid. The words are real. The crossings hold.

But the clues still need a human pass. The AI will happily ship “COCOA” clued as “like cola” because letter-fitting is satisfied. I read every puzzle, mark the weak clues, and paste the fixes back into Claude Code, which updates the grid. That step is not optional. It is the creative judgment that the scripts and the model cannot do.

Level design is the same shape of problem. Parts of it are math. Parts of it are creativity. The modes differ in where the line sits.

If you are building a game right now and thinking about where AI fits, the shortest version is this. Use AI for the math. Keep the creative judgment for yourself.

For pre-generated levels, generate many candidates and hand-pick the ones that feel right. For real-time levels, design the building blocks yourself and let AI arrange them under constraints. For player-generated levels, give players a prompt interface and let the AI handle the tool complexity.