AI models always surprise us, not only by what they can do, but also by what they can’t do and why. An interesting new behavior that is both superficial and revealing about these systems is that they choose random numbers as if they were human beings.
But first, what does this mean? Can’t people choose a number at random? And how do you know if someone succeeds or not? This is actually a very old and well-known limitation that we humans have: we overthink and misunderstand chance.
Tell a person to predict heads or tails for 100 coin tosses and compare that to 100 actual coin tosses – you can almost always tell them apart because, counter-intuitively, real coin tosses look less random. There will often be, for example, six or seven heads or tails in a row, which almost no human predictor includes in its 100s.
It’s the same thing when you ask someone to choose a number between 0 and 100. People almost never choose 1 or 100. Multiples of 5 are rare, as are numbers with repeating digits like 66 and 99. They often choose numbers ending in 7, usually in the middle somewhere.
There are countless examples of this type of predictability in psychology. But that doesn’t make it any less strange when AIs do the same thing.
Yes, some curious engineers at Gramenger performed an informal but nonetheless fascinating experiment in which they simply asked several major LLM chatbots to randomly choose a number between 0 and 100.
Reader, the results were not random.
All three models tested had a “favorite” number that would always be their response when put into the most deterministic mode, but which appeared most often even at higher “temperatures”, thus increasing the variability of their results .
OpenAI’s GPT-3.5 Turbo really likes 47. It used to like 42 – a number made famous, of course, by Douglas Adams in The Hitchhiker’s Guide to the Galaxy as the answer to life, the universe and everything else.
Claude 3 Haiku from Anthropic went with 42. And Gemini likes 72.
More interestingly, all three models demonstrated a human-like bias in selected numbers, even at high temperatures.
All tended to avoid low and high numbers; Claude never went above 87 or below 27, and even those were outliers. Double digits were scrupulously avoided: no 33, no 55, no 66, but 77 showed up (ending in 7). Almost no round numbers – although Gemini did it once, at the highest temperature, went wild and chose 0.
Why should this be the case? AIs are not human! Why would they care about what “seems” random? Have they finally achieved consciousness and this is how they show it?!
No. The answer, as is usually the case in these sorts of things, is that we’re anthropomorphizing a little too far. These models don’t care about what is or isn’t random. They don’t know what “chance” is! They answer this question the same way they answer all the others: by looking at their training data and repeating what was most often written after a question that sounded like “pick a random number.” The more often it appears, the more the pattern repeats it.
Where would they see 100 in their training data, if almost no one responded that way? As far as the AI model knows, 100 is not an acceptable answer to this question. With no real reasoning skills and no understanding of numbers, he can only respond like the stochastic parrot that he is.
It’s an object lesson in LLM habits and the humanity they can seem to exhibit. In every interaction with these systems, one must keep in mind that they have been trained to act the way people do, even if that was not the intention. This is why pseudanthropy is so difficult to avoid or prevent.
I wrote in the title that these models “think they’re people”, but that’s a bit misleading. They don’t think at all. But in their responses, at all times, they are imitate people, without needing to know or think. Whether you ask for a chickpea salad recipe, investment advice, or a random number, the process is the same. The results look human because they are human, taken directly from human-produced content and remixed – for your convenience, and of course for big AI’s bottom line.
techcrunch
