Ask six different AI chatbots to pick six lottery numbers from 1 to 49, and you will mostly get the same handful of tickets back.
Researchers tested six language-model configurations with four different English prompts, logging 1,200 requests for six distinct numbers between 1 and 49. Of those, 1,184 produced valid tickets. The models generated only 8 to 93 distinct combinations each, and in every case one single ticket, the modal pick, showed up in 22.5% to 68% of all valid responses. A measure of how spread out the numbers were, effective diversity, came in between 9.9 and 18.0 for the models, versus a 46.6-to-46.7 threshold expected from genuinely random six-of-49 draws at the same sample sizes.
That gap matters because people already lean on chatbots for quick random picks, whether for lottery numbers, raffle entries, or ad hoc sampling in spreadsheets and experiments. Language models are not random-number generators: they predict the statistically likely next token, so they default to a cluster of favorite numbers rather than spreading evenly across the range. Anyone using an LLM output as a stand-in for randomness is more likely to collide with someone else's random pick than they would with a real draw.
The researchers even pulled two archived Polish Lotto draws for comparison, and real lottery balls landed an effective diversity around 41, more than double what any tested chatbot managed. The paper does not pin down why the models behave this way, or whether a different prompt or a higher temperature setting would fix it. For now, if you want six truly random numbers, a tumbling cage still beats a chatbot.