The number 27 puzzle Why does ChatGPT and other AI always choose it?
If you ask ChatGPT to choose a number between 1 and 50, its most common answer is 27. This may seem like a coincidence, but it's something many experts in artificial intelligence and computational neuroscience are investigating, as it reveals information about the ability of these AI tools to generate randomness.
According to IFL Science, the frequency of this number isn't limited to a specific system. Chatbots from OpenAI, Google, Microsoft, and Antropic frequently produce this result. When asked why they chose it, they said they were looking for a number that seemed random but wasn't typical, such as a multiple of five or ten. Numbers like 17 or 37 are associated with common human choices, so AI ignores them.
They explain that long language models (LLMs) do not understand numbers per se, but rather as symbols, focusing on their properties rather than their mathematical meaning. That is, they are vector-oriented, which limits their ability to generate random choices and favors the emergence of repetition.
Daniel Kang, a professor at the University of Illinois at Urbana-Champaign, says one reason for the repetition may be reinforcement learning from human feedback (RLHF). Training models with answers users like ultimately results in a more appealing answer for the machines, but this isn't the machines' fault alone.
We say this because a study of 200,000 participants revealed that when participants were asked to choose a number between 1 and 100, the most frequently occurring numbers were 7, 37, and 77. This suggests that both humans and machines perceive some numbers as more random than others, although Andrei Karpati, a leading figure in the field of artificial intelligence, asserts that this phenomenon is another example of how language models provide very similar answers.
Other theories, such as that of businessman Chester Zelaya, suggest that the models would employ strategies close to game theory, such as using a binary tree to partition the range of numbers, but this seems unlikely. These theories suggest that these language models operate using probabilities associated with sequences of text, rather than through this type of reasoning, unless explicitly indicated.