AI chatbots narrow knowledge diversity, study of 27 LLMs finds

Researchers at the University of Copenhagen's DIKU tested 27 LLMs across 155 topics and 200 prompt formulations, generating 70 million responses. Even GPT-5's answers were 18.5% less diverse than Google Search results, raising fears of a 'vicious cycle' where popular content dominates.

• 27 LLMs tested on 155 topics and 200 prompt formulations
• 70 million model responses generated
• GPT-5 answers 18.5% less diverse than Google Search
• Researchers warn of a 'vicious cycle' favoring dominant content