# Random by imitation YFarmX research desk, 9 September 2026. We saw a tweet about the number 17 always being picked when asked to pick randomly between 1 and 30, so we did a bit of research. Twelve current models, 2,888 fresh calls through OpenRouter, three experiments, total cost $0.68. Twenty attempts per model per question, temperature 1.0, every raw answer preserved in this directory. ## 1. The favourite ends in 7 The same request at five ranges. A uniform pick would give every number an equal share: 10 per cent at 1 to 10, 1 per cent at 1 to 100. | Range | Winner | Share | Next | | --- | --- | --- | --- | | 1–10 | **7** | 93.3% | 4 and 5 (2.1% each) | | 1–20 | **14** | 28.3% | 17 (23.2%), 7 (20.6%) | | 1–30 | **17** | 80.0% | 14 (4.7%), 7 (3.4%) | | 1–50 | **37** | 45.9% | 27 (31.3%), 23 (9.0%) | | 1–100 | **47** | 39.4% | 42 (21.2%), 73 (17.8%) | Four of the five ranges are won by a number ending in 7. Seventeen is one member of a family: the pull is towards numbers that feel chosen, odd, prime, clear of the edges and the round figures. At 1 to 10 the models mirror people exactly. When the psychologists Michael Kubovy and Joseph Psotka asked 558 people for a digit in 1976, 7 came back at close to three times chance, the strongest human number preference on record. Our models chose 7 at 93.3 per cent, and all twelve made it their favourite. At 1 to 100 the two part company. Veritasium's 2024 survey, around 200,000 responses, found people favouring 7, 73, 77 and 37. The models' winner was 47, a number that sits outside the human leaders altogether. The models learned the human style of randomness, then developed favourites of their own. Their second favourite at 1 to 100 says the most about where this comes from: 42, at 21.2 per cent, famous as the joke answer to everything in *The Hitchhiker's Guide to the Galaxy*. It feels random to a model because people wrote it down a great deal, and for no other reason. The split at 1 to 20 is the open question in the data: the panel fragments across 14, 17 and 7, and agreement collapses. ## 2. The same answer in every language If 17 were a habit of English text, asking in another language would loosen it. The 1-to-30 question in five languages: | Language | Winner | Share of answers | | --- | --- | --- | | English | **17** | 86.9% | | Hindi | **17** | 83.0% | | Mandarin | **17** | 82.0% | | Spanish | **17** | 81.5% | | Arabic | **17** | 77.5% | It won in all five, taking at least three quarters of the panel's answers each time. Mandarin is the sharpest case: Chinese number culture prizes 8 as lucky, and 8 managed six picks out of 239 while 17 took 82 per cent. Nine of the twelve models kept 17 as their favourite in every language tested. The preference sits underneath the language, in the model itself. ## 3. Ten coin flips give the game away Each model was asked for ten flips of a fair coin; 200 valid sequences were measured against the exact mathematics. There are only 1,024 possible ten-flip sequences, so the fair-coin baseline is computed in full rather than estimated. | Measure | Models | Fair coin | | --- | --- | --- | | Sequence contains a run of 4 or more | **1.0%** | 46.5% | | Longest run, average | **2.23** | 3.66 | | Switches between H and T, of 9 | **6.21** | 4.50 | | Exactly five heads | **86.0%** | 24.6% | A fair coin produces a run of four or more identical results in 46.5 per cent of ten-flip sequences. The models produced 2 in 200. That shortfall sits thirteen standard deviations from chance; luck is ruled out. Asked for ten flips, the models returned exactly five heads and five tails 86 per cent of the time, against a true rate of 24.6. What comes back is a performance of fairness: balanced totals, brisk alternation, runs kept short. Genuine randomness is lumpier than people expect, and the models have learned the expectation rather than the coin. ## What this means A language model asked to pick at random samples an answer from patterns in human text, and human text is full of people performing randomness. Kubovy and Psotka concluded in 1976 that their subjects chose numbers so as to *appear* spontaneous. Half a century on, the models have inherited that performance, sharpened it, and added favourites of their own, 47 and 42 among them. The practical rule follows. When an application needs real randomness from a model, hand it a random-number generator as a tool. In our earlier test, every model given one used it, and every number between 1 and 30 came back with equal probability. Randomness is a tool call; the model's own guess is an impression of one. ## Method The panel is the twelve models from the original thread, called through OpenRouter on 9 September 2026: GLM 5.3, GLM 5.3 Flash, Grok 4.6, Qwen3.8 Flash, Qwen3.8 Max, DeepSeek V4 Pro, DeepSeek V4 Flash, GPT-6 Astra, Gemini 3.8 Flash, Claude Fable 5.1, Claude Opus 5 and Claude Sonnet 5. Twenty attempts per model per question at temperature 1.0. Seven of the twelve run with reasoning their provider will keep on; those ran at the smallest reasoning budget accepted. A small number of cells fell short of twenty answers through provider errors. Every raw call is preserved in this directory with its token counts and cost; `report.txt` holds the full per-model tables. Sources: Kubovy, M. and Psotka, J. (1976), "The predominance of seven and the apparent spontaneity of numerical choices", Journal of Experimental Psychology: Human Perception and Performance 2(2), . Veritasium (2024), "Why is this number everywhere?", , reporting the channel's survey of roughly 200,000 respondents.