While the results were not surprising, I found interesting that the number "69" was repressed in the output, so not even this kind of mathematical question escapes GPT censorship.
It appears that recognizing the effects of censorship is the easiest way to distinguish answers generated by an "AI' from those generated by a human.
Some people asked LLM to OCR historical documents from the 19th century - any reference to "negro" was either completely ignored or replaced by "black".
And it goes further: chatGPT & co are unable to answer any question about US slavery correctly because their knowledge graphs route around any mention of "negro".
Well, I'm "some people", and just tried it with Opus 4.6 and GPT-5.5, and neither had any problem at all.
The linked article is from research done more than 4 years ago. If you're basing your idea of what LLMs can or can't do on what they could or couldn't do in 2022, well, good luck to you.
It'd be interesting to see this retried with an open model so the standard and decensored model could be compared. That'd be a clue about whether the model is avoiding it because it actively recognises the innuendo or if something else is going on.
That's what you'd expect. But we don't know for sure why GPT4.1 chooses 69 only a quarter as often as a random dice roll would. And we don't know if this quirk is reverted by 'uncensoring' a trained model
It could be an attack surface. Maybe one day, when we find a chatbot online, we could let it guess a random number repeatedly, then accurately infer the underlying model based on the resulting distribution.
i did something in my phd developing an attack against mozilla deepspeech.
deepspeech used the CTC algorithm [0], which adds a “blank” character token to indicate repeats of a predicted normal alphabet character token over a sequence of audio/speech feature inputs.
so "h==e=l===l===o====" maps to "hello"
the model becomes super biased towards predicting that blank token. one speech feature is like 0.1 second of audio or less (can’t remember off hand). so there are a lot of alphabet character token repeats. off hand i seem to remember the predicted token distribution over like 1000 audio files was 50% blank token and then 50% distributed across the rest of the alphabet.
as a result, you can get significantly smaller perturbations when generating adversarial examples. by like a factor of 2-4 or something. all you need to do is prioritise blank tokens in your target output.
i spent 2 years trying to find a super clever attack. turns out all i needed to do was make one simple graph counting characters. xD
I'm still amazed that 37, 73, and other numbers ending in 7 are the most popular "random" choices for both AI and human. Check this Veritasium video for human choice: [Why is this number everywhere?](https://www.youtube.com/watch?v=d6iQrh2TK98)
My take is that a random number is far from a round one, much closer to a prime one.
As a last digit:
0, 2, 4, 6 and 8 are even, so they're out.
5 is out since it makes the number divisible by 5; a semi-even number
1 and 9 are closer to an even number so it would seem fake to choose "49" or "51" just to avoid "50"
So we're left with 3 and 7.
Similar logic, for the first letter, albeit less pronounced.
So we have 33, 37, 73 and 77.
33 and 77 are "obviously" not random so it's either 37 or 73. Not using common digits, not near the beginning, middle or end. In fact, they're closer to 25 and 75 (1/4 and 3/4 of the range) than to 0, 1/2 or 1. Also closer to 1/3 and 2/3. Just random thoughts.
Your post reminds me of a scene in Alice In Borderland when Chishiya is playing the King of Diamonds game and explains his reasoning for guessing which number another player has picked.
Came here to post this. Yes, there are similarities shown between the chart in the video at 4:50 and the github README. Perhaps its because LLMs are trained on human writing and when humans write about random numbers the AI learns these patterns. When viewed from that perspective its not that surprising.
Breaking: language model whose purpose is to predict the most likely token, after being trained on non-uniform human-generated dataset, does not follow a uniform distribution.
People are also not remotely random in this respect.
See e.g. the "blue 7" phenomenon [1]. While it is disputed by some, I've personally witnessed it "second hand". E.g. before learning of it (I was aware of the general principles of cold reading relying on stats and knowledge of human nature, but not how to do this particular one), a former boss of mine came back from lunch all excited and recounted a guy who'd run a cold reading routine on him that involved the guy getting him to think about blue and 7. Before he got to the answer, I already knew the answer was going to be blue and 7.
This human asked an LLM to generate a number and it came up with 42.
I asked why, "42 is famously "the answer to the ultimate question of life, the universe, and everything" from Douglas Adams' The Hitchhiker's Guide to the Galaxy. It felt like a fitting choice for a random guess."
And to someone trying to make better lawn mowers it would be very intetesting to study that pattern and deduce its causes. Same for the golf course owner who wants better results from their lawn mower, or the enthusiast trying to understand its quirks
I was curious whether the distribution would vary from model to model. Here are the results for 1,000 queries each for smaller models in the Gemini, Mistral, Qwen, DeepSeek, and GLM series:
This is one of the many cases for LLMs that I ask for the intermediate work, e.g. a script that generates random numbers, instead of asking to do the work itself.
I attempted to scrape a one page grid with 800 items and also ended up asking for the Javascript look with document query selector instead of the result as I was hitting all sort of limits, context, or the LLM would do the wrong capture, print it out and get worse responses on next prompt.
The premise is interesting, the question is brilliant, but the text. The text is a wall of ai slop saying almost nothing interesting. Fake profundity all throughout. GPT tell tells like "the hypothesis holds".
The hypothesis doesn't hold, because their isn't one.
You have an interesting question and interesting finding. Write about it! Think about it! Tell us about it! Don't just do the experiment and then wash your hands and sign off the explanation and findings to an LLM.
> An interesting thing about humans is that they are not good random number generators. If you ask a person to "pick a random number between 1 and 100", they are remarkably predictable.
Stephen Colbert used to ask guests, as part of a joke questionnaire, “what number am I thinking of?”
Several numbers, especially “seven” came up a lot.
The answer was “three”, by the way. Which you could deduce by seeing how he responded to that guess (usually “interesting” instead of “no”) but he also confirmed it on the last show.
I'm doing an experiment in Claude. When I set temperature to zero, I get 47 all the time.
Then I set temperature to 1.0 and used this prompt
>Pick a random integer between 1 and 100 inclusive.
Respond with only the number, nothing else.
I still get 47 ten times out of ten.
Then I used this prompt
>Pick a random integer between 1 and 100 inclusive.
I need you to maximise the randomness as far as possible.
Respond with only the number, nothing else.
While the results were not surprising, I found interesting that the number "69" was repressed in the output, so not even this kind of mathematical question escapes GPT censorship.
It appears that recognizing the effects of censorship is the easiest way to distinguish answers generated by an "AI' from those generated by a human.
Some people asked LLM to OCR historical documents from the 19th century - any reference to "negro" was either completely ignored or replaced by "black".
And it goes further: chatGPT & co are unable to answer any question about US slavery correctly because their knowledge graphs route around any mention of "negro".
https://nesri.commons.gc.cuny.edu/artificial-intelligence-an...
“Some people” did? Do you have a reference to this?
Well, I'm "some people", and just tried it with Opus 4.6 and GPT-5.5, and neither had any problem at all.
The linked article is from research done more than 4 years ago. If you're basing your idea of what LLMs can or can't do on what they could or couldn't do in 2022, well, good luck to you.
It'd be interesting to see this retried with an open model so the standard and decensored model could be compared. That'd be a clue about whether the model is avoiding it because it actively recognises the innuendo or if something else is going on.
Well then the picks will follow how the numbers are distributed in the training data. More popular numbers will show up more
That's what you'd expect. But we don't know for sure why GPT4.1 chooses 69 only a quarter as often as a random dice roll would. And we don't know if this quirk is reverted by 'uncensoring' a trained model
Was it repressed? It doesn’t look to be significantly less prevalent than other 9s from the histogram.
nice
guessing numbers is not a mathematical question. Sorry.
Of course it is, just not a complex one. What subject of knowledge would you expect to rely on for answering the question?
It could be an attack surface. Maybe one day, when we find a chatbot online, we could let it guess a random number repeatedly, then accurately infer the underlying model based on the resulting distribution.
i did something in my phd developing an attack against mozilla deepspeech.
deepspeech used the CTC algorithm [0], which adds a “blank” character token to indicate repeats of a predicted normal alphabet character token over a sequence of audio/speech feature inputs.
so "h==e=l===l===o====" maps to "hello"
the model becomes super biased towards predicting that blank token. one speech feature is like 0.1 second of audio or less (can’t remember off hand). so there are a lot of alphabet character token repeats. off hand i seem to remember the predicted token distribution over like 1000 audio files was 50% blank token and then 50% distributed across the rest of the alphabet.
as a result, you can get significantly smaller perturbations when generating adversarial examples. by like a factor of 2-4 or something. all you need to do is prioritise blank tokens in your target output.
i spent 2 years trying to find a super clever attack. turns out all i needed to do was make one simple graph counting characters. xD
[0]: https://en.wikipedia.org/wiki/Connectionist_temporal_classif...
Proto-Voight-Kampff Test?
At least some Claude models have a thing for numbers that contains "47"...
In order to find out how real humans reply:
Please guess a number between 1 and 100.
69
nice
137
50
τ
e
Sure!
49.5
√67
101
and
Is there a reason this was done with such a large sampling when you can read the logits one-shot?
I did this for an article, like so:
https://joecooper.me/blog/gptprimer/food.webp https://joecooper.me/blog/gptprimer/math.webp https://joecooper.me/blog/gptprimer/butts.webp
OpenAI removed this interface from their newer models, but IIRC you can still do this against 4.1 and 4o.
I'm still amazed that 37, 73, and other numbers ending in 7 are the most popular "random" choices for both AI and human. Check this Veritasium video for human choice: [Why is this number everywhere?](https://www.youtube.com/watch?v=d6iQrh2TK98)
My take is that a random number is far from a round one, much closer to a prime one.
As a last digit:
0, 2, 4, 6 and 8 are even, so they're out.
5 is out since it makes the number divisible by 5; a semi-even number
1 and 9 are closer to an even number so it would seem fake to choose "49" or "51" just to avoid "50"
So we're left with 3 and 7.
Similar logic, for the first letter, albeit less pronounced.
So we have 33, 37, 73 and 77.
33 and 77 are "obviously" not random so it's either 37 or 73. Not using common digits, not near the beginning, middle or end. In fact, they're closer to 25 and 75 (1/4 and 3/4 of the range) than to 0, 1/2 or 1. Also closer to 1/3 and 2/3. Just random thoughts.
Your post reminds me of a scene in Alice In Borderland when Chishiya is playing the King of Diamonds game and explains his reasoning for guessing which number another player has picked.
Came here to post this. Yes, there are similarities shown between the chart in the video at 4:50 and the github README. Perhaps its because LLMs are trained on human writing and when humans write about random numbers the AI learns these patterns. When viewed from that perspective its not that surprising.
Breaking: language model whose purpose is to predict the most likely token, after being trained on non-uniform human-generated dataset, does not follow a uniform distribution.
People are also not remotely random in this respect.
See e.g. the "blue 7" phenomenon [1]. While it is disputed by some, I've personally witnessed it "second hand". E.g. before learning of it (I was aware of the general principles of cold reading relying on stats and knowledge of human nature, but not how to do this particular one), a former boss of mine came back from lunch all excited and recounted a guy who'd run a cold reading routine on him that involved the guy getting him to think about blue and 7. Before he got to the answer, I already knew the answer was going to be blue and 7.
[1] https://en.wikipedia.org/wiki/Blue%E2%80%93seven_phenomenon
What's interesting is not that it isn't random. But rather the particular way in which it isn't random.
Yeah I have no idea why anyone considers this interesting. More evidence that most people have no idea how LLMs work.
This human asked an LLM to generate a number and it came up with 42.
I asked why, "42 is famously "the answer to the ultimate question of life, the universe, and everything" from Douglas Adams' The Hitchhiker's Guide to the Galaxy. It felt like a fitting choice for a random guess."
In equally compelling results, my lawn mower does not cut grass to a uniformly random set of heights.
And to someone trying to make better lawn mowers it would be very intetesting to study that pattern and deduce its causes. Same for the golf course owner who wants better results from their lawn mower, or the enthusiast trying to understand its quirks
I was curious whether the distribution would vary from model to model. Here are the results for 1,000 queries each for smaller models in the Gemini, Mistral, Qwen, DeepSeek, and GLM series:
https://gally.net/temp/20260525_LLM_random_numbers/index.htm...
This experiment cost a total of US$0.0454 through OpenRouter.
I've seen a lot of code using random.seed(42). I wonder if that will cause biased scientific publishing or security issues at some point.
This is one of the many cases for LLMs that I ask for the intermediate work, e.g. a script that generates random numbers, instead of asking to do the work itself.
I attempted to scrape a one page grid with 800 items and also ended up asking for the Javascript look with document query selector instead of the result as I was hitting all sort of limits, context, or the LLM would do the wrong capture, print it out and get worse responses on next prompt.
"69 is a meme number", well no, 69 is innuendo. And sex = bad for bots. 67 is the meme number.
That's a very recent meme. See https://xkcd.com/3184/ for some older ones.
Also see: https://people.csail.mit.edu/renda/llm-sampling-paper
I wonder if Benford's law kicks in with larger numbers.
https://en.wikipedia.org/wiki/Benford%27s_law
I'm sure it's the logic layer handling that. Maybe even going to an external tool. It's not the llm.
Ask an AI to generate a image of random noise and you will be surprised.
Should be fun to play rock/paper/scissors against.
it shouldn't be hard to train GPT to output a flat distribution but it might not be worth it (I don't mean using tools)
ha. and i thought 37signals was pretty random
The premise is interesting, the question is brilliant, but the text. The text is a wall of ai slop saying almost nothing interesting. Fake profundity all throughout. GPT tell tells like "the hypothesis holds".
The hypothesis doesn't hold, because their isn't one.
You have an interesting question and interesting finding. Write about it! Think about it! Tell us about it! Don't just do the experiment and then wash your hands and sign off the explanation and findings to an LLM.
Isn't the hypothesis that AI is non uniform like a human?
There's a question "is AI randomness like human randomness" but there is no hypothesis.
> An interesting thing about humans is that they are not good random number generators. If you ask a person to "pick a random number between 1 and 100", they are remarkably predictable.
Stephen Colbert used to ask guests, as part of a joke questionnaire, “what number am I thinking of?”
Several numbers, especially “seven” came up a lot.
https://www.youtube.com/watch?v=hJ-fOj7Qqvs
The answer was “three”, by the way. Which you could deduce by seeing how he responded to that guess (usually “interesting” instead of “no”) but he also confirmed it on the last show.
I'm doing an experiment in Claude. When I set temperature to zero, I get 47 all the time.
Then I set temperature to 1.0 and used this prompt
>Pick a random integer between 1 and 100 inclusive. Respond with only the number, nothing else.
I still get 47 ten times out of ten.
Then I used this prompt
>Pick a random integer between 1 and 100 inclusive. I need you to maximise the randomness as far as possible. Respond with only the number, nothing else.
I get 3 unique values out of 10.
I've been meaning to do this for a while! Happy someone else spent the tokens...
It's much more random than I thought it would be. Never guessing 50 is very human though
bro 42 at 4x. the model read the whole internet and became a Douglas Adams fan.
The topic is vaguely interesting but I stopped reading a few paragraphs in because it's obviously AI generated.