The phenomenon: asking for a random number → answer “47”
Perhaps this has happened to you: you open ChatGPT and say something like:
“Tell me a random number from 1 to 100”
And it gives you back: “47”.
Other people report that the number is repeated, that it doesn't seem so "random", that it becomes recurrent. On forums like Reddit This phenomenon is discussed:
“Every time you ask ChatGPT to pick a number between 1 and 50, it picks 27.”
“"This time I asked why? 47 said it was the most random number between 0 and 100."”
Although not all comments mention 47 — some mention 27, 37, or others — the pattern is that the model tends to repeat certain numbers instead of perfectly distributing the choice.
What's really going on?
To understand this, it's helpful to remember how ChatGPT (and large language models in general) works and why it "chooses" a number that appears random but is actually biased. Some key points:
- There is no pure internal “random number generator”
When you ask for “a random number”, ChatGPT is not running a uniform random number generation procedure in the style ofrand(1,100). Instead, it's generating predicted text — given the prompt words, what word (or number) is most likely to follow based on its training. - Prediction based on sequence probability
Instead of "choosing" numbers from 1 to 100 with equal probability, the model is determining which number "sounds" more plausible or more frequent in its training corpus, or "more likely" as a text response to that prompt. Therefore, certain numbers may have "more weight" because they appear more often or have stronger linguistic associations. - Human expectancy effect and linguistic biases
When you ask for “a random number,” you also force a (human) expectation of “something that doesn’t seem calculated.” The model can learn (or reflect) that certain common “random” responses are more socially acceptable or more “surprising.” For example, 47 is odd, non-trivial (not 50, not 100, not 1), and may have been more frequent in training examples, conversations, memes, etc. - Prompt variability, temperature, context
Depending on the exact prompt (“Give me a random number from 1 to 100”), the system (model version), the internal “temperature” or randomization, and the conversation history, the response may change. But in many cases, the “popularity” of certain numbers is observed, leading to the impression that the same number is always chosen.
Why the number 47 in particular?
There's no official confirmation that ChatGPT has a fixed predisposition to the number 47. However, the Reddit community notes that the number 47 appeared frequently in their experiments. Some reasoning follows:
- It is a number that is “random enough” for humans: neither too low nor too high, not an obvious multiple, “not too common” like 50 or 10.
- There could be a “selection bias” in the data: for example, if many “pick a random number” examples on the internet used 47, then the model “learned” it as plausible.
- The text generation structure predicts what is "most likely" given the preceding words; if in your training "47" appears more frequently than other numbers in that context, it will tend to choose it.
That's how one sums it up Reddit comment:
“"It's just generating the most random sounding number. Same happens when humans make up random numbers." (Reddit)
And another:
“Pretty sure this came up… It is the internet 😀… AI always picks 27!”
So: there is no hidden magic in 47, but a reflection of how language models work.
Reflection: What does this teach us about the fundamentals of language and “randomness”?
Apparent randomness vs. true randomness
When we ask for a “random” number, we expect all values from 1 to 100 to have the same probability of being selected. But with ChatGPT, the “randomness” is simulated: text is generated based on conditional probabilities, not a uniform number generator. This leads us to ask: what does it mean for something to “appear random” but not actually be?
In many fields—writing, human language, creativity—what we consider “random” isn’t truly uniform, but rather what “doesn’t seem to follow a recognizable pattern.” And that’s enough in many contexts. Language models, by predicting what is most likely, achieve this appearance without designing true uniformity.
Learning, frequency, and bias
The fact that a model frequently repeats the number 47 shows that training patterns matter. If a model is trained on corpora where certain responses are common, then those responses will replicate the structure of those corpora. This reminds us that models don't "understand" in the human sense, but rather learn text patterns. And those patterns can lead to unexpected biases (such as preferring the number 47).
The basis of language prediction
This example illustrates in a simple way how a language model operates:
- Receive a prompt (“Tell me a number from 1 to 100”).
- Based on your training, evaluate which number is most likely to follow the prompt.
- Give me the number.
Although it may seem that he is "choosing" freely, he is actually applying what he has learned as a "plausible response".
Lessons for AI users and developers
- As a user: if you expect true randomness, it's best to use a random number generator (e.g.,
random()in code), and not depend on a language model for that. - As a developer or analyst: understand that models like ChatGPT have implicit limitations when asking for "chance", "randomness", "totally free creativity": what they offer is linguistic probability, not genuine randomness or free thinking.
- In general: if we see that a model frequently repeats certain responses (“47”, “27”), it is a sign that the training data or architecture is having a strong influence, and that what is presented as “random” may contain detectable patterns.
A final thought: the role of human perception
Ultimately, this situation also reveals something about us: humans seek out the strange, the random, the unexpected, and when something repeats a pattern (like "he says 47 again"), we detect and question it. This leads us to place more importance on the fact that "47 again" than on the fact that, statistically, there might not be a pattern at all.
And that human awareness—of pattern, of surprise, of randomness—is what differentiates the user's experience from the internal mechanics of the model. Models generate, we interpret.
Conclusion
Yes, when you ask ChatGPT for a number between 1 and 100, it might respond with "47" (or a number that appears more often than you'd expect). But this isn't due to a hidden trick: it's simply the nature of the language model, its training with text patterns, and its prediction mechanism.
This simple experiment connects us to broader questions: What is “random”? How does human and artificial language operate? To what extent do we perceive patterns that may not exist or may be inherent to the system?