Large language models from major providers display surprising predictability in their responses, according to an investigation by MIT Technology Review. When asked to generate a random number between 1 and 10, chatbots including Claude, ChatGPT, and Gemini consistently return 7, followed by 3 or 4 on subsequent requests, then 8 or 9. While this pattern doesn’t occur every time, the frequency suggests a deeper issue with how these models generate supposedly random outputs.
This predictability reflects what researchers describe as a “groupthink groove” affecting current LLMs. The phenomenon highlights fundamental limitations in how these systems produce varied responses, even when randomness would be expected. According to MIT Technology Review, an unnamed startup is now working to address this groupthink problem, though the publication did not provide specific details about the company’s approach or solutions.
The groupthink issue extends beyond simple number generation, potentially affecting the diversity of ideas and solutions these AI systems can produce across various applications. The consistent patterns in LLM outputs raise questions about the extent to which these models can truly generate novel or unexpected responses when needed.