🤖 GPT-4.1’s random numbers aren’t random

An enthusiast named Exmergo asked GPT-4.1 to pick a random number from 1 to 100, 10,000 times. The results failed a chi-square test with a value of 15,604 and p near zero, showing the choices are far from truly random.

GPT-4.1 avoided round multiples of ten except for 10 once, and nearly ignored the extremes 1 and 100. The number 42 appeared four times more often than expected, a nod to “The Hitchhiker’s Guide to the Galaxy.” Numbers with 7 were favored, while 69 appeared much less, likely due to training filters.

The model reflects human biases in number selection despite being asked for randomness. It seems to have absorbed our quirks from the texts it was trained on.

📊@tech