An experiment tested 65 AI models from 30 companies with 127,918 prompts, asking everyday questions ten times per model without extra context. Answers often converged: 73% of book recommendations named Man’s Search for Meaning, 76% of business ideas were subscription boxes, and common suggestions included Lisbon for a European trip and leather wallets as gifts. Convergence was strongest in consequential advice: 98% recommended a broad index fund for $10,000, 99% named Bitcoin as the cryptocurrency to buy, and 90% suggested computer science to study. Models often underestimated how repetitive they were. Adding a brief personal description shifted recommendations into new patterns, but people with similar descriptions tended to receive the same answer.
The sameness was largely hidden by conversational variety: asking for another option within one chat produced different books, while starting fresh chats repeatedly returned to the same default. Models also invented highly stereotyped people: 69% of 3,151 fictional adults were exactly 34, and particular companies favored recurring names. Harmless answers such as colors, cities, and names could often reveal which company made a model. Tests on an open model traced a favored answer to the stage when it was trained to act as an assistant, rather than simply to information copied from the internet. The findings suggest that fluent, personalized-sounding advice can reflect shared defaults rather than individual insight. To get more relevant options, provide specific personal context or request a list and choose for yourself; simply asking an AI to be original is unlikely to help.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.