AI models continually surprise us not only with what they can do, but also with what they can't do and why. One interesting new behavior about these systems is both superficial and thought-provoking: AI models choosing random numbers like humans do.
But first, what does that even mean? Can't people just pick numbers randomly? And how do we know if someone is doing a good job at it? This is actually a very old and well-known limitation of us humans: we overthink and misunderstand randomness.
If you ask a person to predict whether a coin will flip 100 times and compare that to flipping a real coin 100 times, you will be able to distinguish them most of the time because, counterintuitively, real coin flips don't seem all that random. For example, it's common to get six or seven heads or tails in a row, but a human predictor would rarely include them in the 100.
The same is true if you ask people to pick a number between 0 and 100. Almost no one picks 1 or 100. Multiples of 5 and repeating numbers like 66 and 99 are rare. Most people pick numbers somewhere in the middle, ending in 7.
Psychology is filled with examples of this kind of predictability, but when an AI does the same thing it's no less strange.
Yes, inquisitive engineers at Gramener conducted an informal but interesting experiment: they asked some of their leading LLM chatbots to randomly select a number between 0 and 100.
The results, dear reader, were not random.
Image credit: Gramener
The three models tested had a “favourite” number that was always the answer when put into the most deterministic mode, but this number also appeared most frequently at higher “temperatures”, increasing the variability of the results.
OpenAI's GPT-3.5 Turbo actually prefers 47; previously it preferred 42, which is, of course, the number made famous by Douglas Adams in The Hitchhiker's Guide to the Galaxy as the answer to Life, the Universe, and Everything.
Claude 3 Haiku from Anthropic was 42. And Gemini likes 72.
What's even more interesting is that all three models showed human-like biases in the values they selected, even at higher temperatures.
Everyone tended to avoid low and high numbers, and even those were the exception, with Claude never choosing a number over 87 or below 27. Double digit numbers were strictly avoided, there was no 33, 55 or 66, although 77 (ending in 7) did appear. Fractions were few and far between, although Gemini did go off the rails once, during the hottest hour, and choose 0.
Why would this be? AI is not human! Why would it care about something that appears random? Could it be that the AI has finally gained consciousness and is expressing it this way?
No. The answer, as is often the case with these things, is that we've taken anthropomorphism a step too far. These models don't care what is random and what isn't. They don't know what “randomness” is. The answer to this question is the same as the answer to every other question: look at your training data and repeat what was most often written after a question like “pick a random number.” The more often it appears, the more often the model will repeat it.
Almost no one answers 100, so where is the number 100 in your training data? As far as the AI model knows, 100 is not an acceptable answer to that question. With no real reasoning ability and no understanding of numbers whatsoever, it can only answer like a probabilistic parrot.
This is an example of the habits of LLMs and the humanity they appear to exhibit. When interacting with these systems, we must keep in mind that they are trained to behave in a human-like manner, even if it is unintentional. This is why pseudo-humanity is difficult to avoid or prevent.
In the headline, we said that these models “think they're humans,” but that's a bit misleading. They don't think at all. But their responses always mimic humans, without them needing to know or think. Whether you ask them for a chickpea salad recipe, investment advice, or random numbers, the process is the same. The results feel human because they're made by humans, drawn directly from human-created content and remixed for user convenience and, of course, the benefit of Big AI.