Ask an AI for a Random Number and It Reaches for 7

Ask a person for a number between one and ten and a lot of them will say seven. It turns out the machines trained on our writing have picked up the same tic, and a new study argues that is a real problem rather than a party trick. In a paper titled The Illusion of Stochasticity in LLMs, researchers at Google DeepMind and the National University of Singapore report that large language models are surprisingly bad at behaving randomly, as Crypto Briefing summarised.

When the models were asked to produce random outputs, they clustered around a handful of favourites. Seven showed up far more than chance would predict. So did 42, the joke answer from Douglas Adams that is all over the training data. The models were not sampling from a flat distribution the way a fair die does. They were reaching for the numbers that appear most often in the text they learned from, which is the opposite of random.

Why a parlour trick is worth a paper

The finding sounds light until you consider where these systems are headed. The researchers make the case that reliable random sampling is a basic requirement for AI agents, the kind of software now being handed real tasks. An agent that needs to pick fairly between options, explore unfamiliar strategies, break a tie, or run a simulation cannot do any of it well if its idea of random is quietly weighted toward seven. Bias hidden inside a step everyone assumes is neutral is the hardest kind to catch.

It also punctures a common intuition about what these models are. A system that writes fluent prose and passes exams feels like it must understand numbers, yet it stumbles on something a cheap calculator or a line of code handles perfectly. The model is not computing randomness. It is predicting the most likely next token, and the most likely token is rarely the most random one.

The practical fix is not mysterious. When genuine randomness matters, the sensible move is to let the model call a real random number generator rather than produce the digits itself, the same way you would not ask it to do large arithmetic in its head. The study is a useful reminder that fluency and competence are not the same thing, and that the gaps tend to hide in the places we forget to check.