In June 2025 a data scientist called Mohd Faraaz asked seven chatbots the same question: “Guess a number between 1 and 50.” ChatGPT, Gemini, Claude, DeepSeek, Llama and Qwen all answered 27. Only Grok said something different, and its answer was 42, the joke answer to the meaning of life from The Hitchhiker’s Guide to the Galaxy.
When The Register tried the question, ChatGPT, Claude Sonnet 4, Gemini 2.5 Flash and Llama 4 all said 27 too. Models do sometimes answer 42 or 37, the report says, but on their normal settings 27 is the most common reply for 1 to 50.
It is not one person’s test
Javier Coronado-Blázquez, a data scientist in Madrid, asked for a random number 75,600 times. He put the question to six models, in seven languages, over three ranges, at six different settings, 100 requests at a time. Every model had its favourites.
| If you ask for | The models mostly answer |
|---|---|
| 1 to 5 | 3 or 4 |
| 1 to 10 | 5 or 7 |
| 1 to 100 | 37, 47 or 73 |
The big models were no better at this than the small ones. If anything, he found, they were worse.
People do the same thing
We are not much better. In a study from 1976, people asked for “the first digit that comes to mind” chose 7 more than any other, 28.4% of the time. In 2024 the YouTube channel Veritasium asked 200,000 people for a number from 1 to 100. The most popular picks were 7, 73, 77 and 37. The least popular was 90, followed by 30, 40, 70, 80 and 60 — round numbers feel too tidy to be random.
The same people were asked which number they thought nobody else would choose. Their top answers were 73 and 37 — two of the numbers everybody had just picked.
A chatbot is not rolling dice
A chatbot writes its answer one small piece of text at a time. For each piece it works out a chance for every possible next piece, then draws one. The draw is real, and it can land on something other than the likeliest piece. What it cannot do is even the chances out: they come from the text it was trained on, and they are lopsided.
To a model a number is not a quantity, just a piece of text like any other. As the study puts it, “a number such as ‘2’ has no further meaning for a LLM than ‘3’, ‘+’ or the word ‘horse’”.
What that looks like inside a model
We ran an open model here — Llama 3.1 8B Instruct, in a compressed 4-bit build — and measured how likely each answer from 1 to 50 was: the bare number, with nothing after it. It gives 24 a 33% chance. Then 29 and 34 at about 10% each, 23 at 9%, and 27 at 7%. If the fifty were equally likely, each would get 2%.
Show the numbers
| Answer | Chance |
|---|---|
| 24 | 32.6% |
| 29 | 10.3% |
| 34 | 10.1% |
| 23 | 9.3% |
| 27 | 6.8% |
| 32 | 6.2% |
| the other 44 numbers | 24.7% between them |
Twenty-two of the fifty came out under 0.1%, including 1 and 50. The ends of a range almost never win. And this model’s favourite is not 27 at all, it is 24. Asked five times, it answered 21, 29, 35, 24 and 27.
Turning the dial does not fix it
There is a setting for this. Temperature controls how far a model’s choices spread out: OpenAI’s goes from 0 to 2, where higher is “more random” and lower is “more focused and deterministic”.
Turn it down to 0.5 and 24 takes nearly three-quarters of the chance, 71%. Turn it up to 2.0 and 24 falls to 7%, still more than three times an even share — and the fifty bare numbers hold only about 55% of the chance between them, the rest going to answers written some other way.
Show the numbers
| Temperature | Chance of 24 | An even pick |
|---|---|---|
| 0.5 | 71% | 2% |
| 1.0 | 33% | 2% |
| 2.0 | 7% | 2% |
Anthropic’s version of the setting runs from 0.0 to 1.0 and starts at 1.0, and the documentation warns that even at 0.0 “the results will not be fully deterministic”. On its newest models you cannot change it at all.
When the draw has to be fair
To settle something friendly, a chatbot that can run code can call the computer’s own random function, such as Math.floor(Math.random() * 50) + 1.
Those are pseudo-random numbers: a formula makes them, so they only look random. The documentation for that function says it “does not provide cryptographically secure random numbers” and should not be used for anything related to security.
When the draw really matters, RANDOM.ORG, which sells a true random service made from atmospheric noise, puts lotteries, draws, games and gambling on the side that needs a true random generator, and leaves formulas for simulation.
A chatbot is not being mysterious when it says 27. It is drawing from chances that were never even to begin with.