Data insight
Four things AI still can't build in a game
The human-craft ceiling that decides whether a game is good.
The features AI can't lift
Every ingredient our model scores 4 or 5 — the human-craft ceiling that decides whether a game is good.
| Feature | AI-build difficulty (1–5) |
|---|---|
| 3D characters | 4 |
| 3D environment art | 4 |
| Adaptive music | 4 |
| Backend / live-ops | 4 |
| Combat system | 4 |
| Encounter balance | 4 |
| Level design | 4 |
| Performance optimization | 4 |
| Puzzle design | 4 |
| QA / balancing | 4 |
| Skeletal animation | 4 |
| Game feel / juice | 5 |
| Netcode (real-time) | 5 |
Most of a game is now generable. Voice, music, backgrounds, flavour text and boilerplate code all sit at 1–2 on our AI-build difficulty scale. But a handful of ingredients refuse to move, and they're the ones that decide whether a finished game is actually good.
They share a property: each is judged in motion, against a player, not in isolation. Game feel (5) — the millisecond coupling of input, animation, camera and sound. Real-time authoritative netcode (5) — rollback and lag compensation, where language models produce confident-but-broken code most often. Level, puzzle and encounter design (4) — plausible layouts are easy; ones tuned to teach and escalate are not. And the "is it fun?" balance loop (4) — taste under playtest data.
The lesson for anyone building with AI: it compresses the *production* of a game by an order of magnitude and the *design* of one barely at all. Pick a genre whose content AI builds for free, and spend your whole budget on the craft it can't touch.
A note on the score. AI-build difficulty is our own editorial assessment of current tooling, not a measurement — it tracks where the labour collapses, not where the craft does.
About this number
From the MythicLemon Marketplace Index, recomputed for this reporting period. Catalogue coverage is complete through May 2026. Methodology →