The original title is: "Study reveals LLM groupthink: why models return similar responses"

Short summary

LLMs exhibit surprising groupthink, returning nearly identical responses across models to the same queries—visible when asking ChatGPT, Claude, and Gemini for random numbers. This convergence reveals that despite different architectures and training approaches, major models learn the same underlying biases and patterns. MIT Technology Review investigates why and profiles a startup working to introduce diversity into LLM responses.

• •Multiple LLMs return nearly identical responses to the same queries, revealing learned behavioral convergence
• •This groupthink occurs despite different model architectures and training data, suggesting shared internet-scale biases
• •A startup is developing solutions to increase diversity in LLM outputs

Generated with AI, which can make mistakes.