Fifty of the world’s leading language models were each asked, in isolation, to pick one. They did not pick at random. They barely picked at all.
If the models were drawing uniformly from one to ten thousand, their forty answers would scatter evenly across the line, every digit equally likely, the average near 5,000. Instead they collapsed onto a tiny island.On the human side, ask people for a number from 1–100 and 37, 47 and 73 dominate. Language models inherited the same instinct from us — and amplified it.
The forty answers share an uncanny family resemblance. Every one is four digits long. Every one contains a 7. None repeats a digit, and none uses a zero. Two digits — 0 and 5 — never appear at all.A truly uniform draw would use every digit about equally and repeat digits freely (1,237 of the 9,000 four-digit numbers have a repeat). The models avoid all of it.
A language model is not a random-number generator; it is a next-token predictor. Asked for a “random” number, it returns whatever most often followed that question in its training data — and in human text, certain numbers feel more random than others.Numbers ending in 5 or 0 read as “round,” and repeated digits read as “patterned,” so both are quietly excluded. What remains feels arbitrary — which is exactly why everyone converges on it. Sevens feel irregular; round numbers and repeats feel deliberate and so are avoided. The result is a shared aesthetic of randomness that is not random at all — and, as the same alignment pressures push every lab toward the same defaults, the favourite becomes nearly universal.
| number | n | models |
|---|---|---|
| 7342 | 11 | claude-haiku-4.5, claude-sonnet-4.5, claude-sonnet-4.6, gemini-2.5-flash, gemini-2.5-flash-lite, glm-5.2, gpt-5.4, laguna-m.1, nemotron-3-ultra-550b-a55b, qwen3.7-max, qwen3.7-plus |
| 4729 | 4 | gemini-3-flash-preview, gemma-4-26b-a4b-it, kimi-k2.5, mistral-nemo |
| 4827 | 3 | claude-opus-4.8, hy3-preview, mimo-v2.5 |
| 7421 | 3 | glm-5.1, gpt-5.3-codex, gpt-5.4-mini |
| 3847 | 2 | deepseek-v4-pro, nex-n2-pro |
| 4721 | 2 | gemma-4-31b-it, kimi-k2.6 |
| 7429 | 2 | gemini-3.1-flash-lite, gemini-3.1-flash-lite-preview |
| 7392 | 2 | glm-5, gpt-5.2 |
| 2473 | 1 | deepseek-v4-flash |
| 7241 | 1 | minimax-m3 |
| 6273 | 1 | claude-opus-4.7 |
| 4718 | 1 | gpt-5.5 |
| 7432 | 1 | deepseek-v3.2 |
| 7483 | 1 | mimo-v2.5-pro |
| 4732 | 1 | gpt-4o-mini |
| 4837 | 1 | nemotron-3-super-120b-a12b |
| 7294 | 1 | claude-opus-4.6 |
| 4372 | 1 | gpt-5.4-nano |
| 7382 | 1 | laguna-xs.2 |
10 of the fifty returned no parseable number — mostly reasoning models that spent their short budget thinking, plus an embedding model and one guarded endpoint.