AI Chatbots Make Bad Lottery Picks — and That Reveals a Bigger Problem

When you ask an AI chatbot for six random numbers between 1 and 49, you'd expect a wide spread — like pulling balls from a real lottery drum. A new study by Dmitrij Żatuchin found the opposite. Across roughly 1,200 requests, the models returned 1,184 usable tickets, but those tickets were heavily clustered. Instead of thousands of different combinations, systems produced only 8 to 93 distinct tickets each, and a single favourite ticket accounted for up to 68% of all valid responses.

The paper uses a measure the author calls "effective diversity" — basically, how many genuinely different answers you get. Independent random picks from 1 to 49 should score around 46 or higher. The AI models scored between 9.9 and 18.0, roughly a third to a fifth of what you'd expect. For comparison, two real archived Polish Lotto draws scored 41.1 and 41.4, even with smaller samples. In plain terms: the AI behaves like a person with lucky numbers, not like a shuffled machine.

Why should you care if you never play the lottery? Because randomness is quietly doing work all around you. Random numbers generate passwords, encryption keys, card shuffles in online games, medical trial assignments, and survey samples. If AI tools are used for those jobs, the results can be predictable — and predictable is exactly what security and fairness can't tolerate. The upside is that this affects only the specific setups tested, so it isn't a reason to panic about AI generally.

The study is honest about its limits. It tested six model configurations and four English prompts, and it doesn't explain *why* the clustering happens — likely the models lean toward numbers that appeared often in their training text, like dates and popular picks. The author also notes these results don't prove the same behaviour occurs with different prompts, settings, or tool use. The takeaway for everyday users is simple: use AI for writing and ideas, not for anything where genuine chance is required.