State of AI game making

79%n = 229 of 288 first projects · machines first opened 20 Aug–27 Sep 2026

Asked for their first game, 229 of 288 people chose 3D. Two cross-checks agree: when the AI summarised the idea it read it as 3D 71% of the time (97 of 137), and first messages that named a dimension said 3D 78% of the time (76 of 98).

The received wisdom runs the other way. 2D is cheaper to build, faster to iterate and easier to make look finished, which is why tutorials start there. People arriving with an AI start from the game they picture, and the game they picture has a camera behind a player controller.

14 wordsn = 261 accounts · 13 Aug–29 Sep 2026

The middle half of first messages runs from 6 to 64 words. 46% are 12 words or fewer, the length of a genre and a setting. At the other end, 19% are over 150 words: a full design document pasted in before the AI has asked anything.

Two very different openings, and neither is a brief. The short one leaves every decision to the AI; the long one makes every decision before any of them has been playtested.

79%n = 442 of 558 answers, 101 machines · 24–29 Sep 2026

When the AI asks a question before it builds, it offers answers to pick from. 79% of replies picked one; 17% wrote their own.

That makes the suggested answers the real design decisions in a first game, which is a responsibility as much as a convenience. The window is six days, and the next edition will say whether the share holds.

3 min 1 sn = 147 of 195 machines that got an answer · machines first opened 20 Aug–27 Sep 2026

Of the people who got a first answer from the AI, 75% went on to press Play, and the median press came 3 minutes 1 second after that answer, after a median of one answer.

People do not wait for a finished game before they run it. They run the first thing that exists.

42 sn = 1,253 runs on 220 machines · 11–29 Sep 2026

The middle half of runs lasts from 15 seconds to 2 minutes 31 seconds. 58% end inside a minute. These are not crashes: one of the 1,253 runs ended early.

A 42-second playtest is long enough to check whether the jump works and not long enough to learn whether the level is any good. It is the loop working as intended: play, stop, say what is wrong, play again.

9 min 59 sn = 195 machines that played in their first sitting · machines first opened 20 Aug–27 Sep 2026

For people who pressed Play in their first sitting, the median time from first opening the app to that press was 9 minutes 59 seconds. The middle half ran from 5 minutes 25 seconds to 17 minutes 8 seconds.

That time includes signing in, describing the game, answering the AI's questions and waiting for the first build.

14%n = 37 of 261 first messages · 13 Aug–29 Sep 2026

14% of first messages name a game that already exists, and 23% of accounts name one at some point (61 of 261). No single title reaches the 20-person floor, so this edition names none.

A reference game is the fastest brief there is, and the most dangerous: it carries a budget nobody said out loud.