AIs have a favorite number, and it's not 42 Asked to guess a number between 1 and 50, OpenAI's ChatGPT, Anthropic's Claude Sonnet 4, Google's Gemini 2.5 Flash, and Meta's Llama 4 all provided the same answer: 27. Those who see conspiracies everywhere might be inclined to see one here, but that's not what's going on. There's no collusion beyond common training data that …
Me: Please name a vegetable
ChatGPT: A vegetable is edible.
Me: Please name another vegetable
ChatGPT: Vegetables are healthy.
Me: Please name one more vegetable
ChatGPT: I'm no vegetarian and my system does not run on vegetables. Please rephrase your question.
...
Running Ollama and the mistral:latest model on my slightly elderly desktop PC, I asked it to guess a number between 1 and 50.. and it came straight back with 27.
Spookily, if you assign values to each of the letters t-w-e-n-t-y-s-e- etc, and add them up, and multiply by a hidden coefficient, then add a predefined offset, you can get 42.
Brrrr. Scary.
...it's how AI tends to work for many things.
It simply outputs something around the middle. It's one of the giveaways for AI images and videos (and I'm guessing music)
If you look properly at AI imagery, it tends to look pretty flat, no deep blacks, no bright whites, no hard shadows and bright spots, just an average all over brightness and contrast, like everything has been shot with all over flood lighting.
"guess a number between 1 and 50" (inclusive?) is not the same as "give me a random number between 1 and 50". 27 might be the most popular choice by humans, so the LLM is trying to predict that based on its training.
Which is quite clever in a way, but I don't need anyone else second-guessing me. I have cow-orkers and manglers who do that already.
Yeah, and 69 is also fairly balanced in its own way imho (1 to 100 range), at least for humans (in couple?) ... gotta wonder what the AI RotM "thinks" about that!
Great points! I can sure see how 96 relates numerologically to that French homemade lunch menu special of "soupe au cul tourné" ... and with the right Saharan scorcher hitting around these dangling parts as we speak, even the best laid plans may indeed need to temporarily cut the hotness in half with a 68 and a promise ...
An AI could never be this perspicacious, imho (still lotsa work to do to get that tech up to that level!)!
"Great" (ahem) minds think alike. I was going to say that if the choice hadn't been restricted to between 0 and 50, it would have been easily guessed... "69, dudes!"
Thing is, they're not random number generators, but it's pretty easy nowadays to get them to call a random number generator as a tool. It's trying to pick the most probable answer, not the most random one.
As they're fundamentally built on language, LLMs tend not to do amazingly well at maths. Tokenisation and probabilistic answers tend not to work too well when there is only one right answer.
The current state of the art GenAI systems are excellent at many things, but they have weaknesses due to the nature of them. If the question you're asking is requiring either proper randomness or is a binary one with either a pure right or wrong answer, then it's like using a wrench as a hammer. It might work, but you're probably better off using the right tool for the job.
*I think models isn't really the right term for them to be honest, as they've got many different models but also control systems, etc...
Claude's quoted explanation for how it 'chose' a number in the range 1 to 50 makes sense.
I believe a human - that is not one harbouring pseudo-random number generators in his head - might reason likewise. The choice is based upon 'typicality', but seeks to avoid the most typical number, i.e. 25/26, based on 25.5 being the nominal median value.
Humans are often asked to pick numbers within specified ranges as part of a gambling game or other amusement. Lottery players have an incentive to select a combination of numbers unique from those of competitors because thereby that, as the winning combination, scoops the pool. For people who understand the mechanics of pseudo-random-number generation, or that of allegedly truly random number generation, there is no such thing as lucky/unlucky individual numbers or combinations. Yet, because other people may regard, say, 13 as 'unlucky', it makes psychological sense to include the number.
People may avoid sequences such as 1,2,3,4,5 in the mistaken belief that these are less likely to arise as winning combinations than seemingly haphazard combinations.
Perhaps someone might care to instruct an AI about the nature of lotteries, ask it to choose sets of, say, 5 integers, from, suppose, the range 1 to 50, and 'explain' each combination?
Right on! And it'd be great if they could also develop some sort of luck virus for that AI, so it picks the right lottery numbers each time ... except in Norway of course, where it don't matter at all really -- they promise 1,000,000kr and then give winners just 100kr ... what a svindel!
What's wrong with choosing eg, 30.56?
Think about it. You ask an "intelligent" person to "guess a number between 1 and 50", how many of those people will (because they want to appear "super-intelligent") pick holes in your question, just to be contrary, and give an answer that answers the question, but does not answer the question that you intended to ask, which is likely to have been "guess an integer between 1 and 50".
So AI is not yet at the point where it is behaving like a smart-arse intelligent person.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
