Study finds GPT-5 answers 18.5% less varied than Google Search

A study by researchers at the University of Copenhagen found that answers from large language models (LLMs) were less varied than information available through Google Search. Even GPT-5, the model that produced the most varied responses in the study, gave answers that were 18.5% less varied than Google Search results.

The researchers tested 27 LLMs across 155 topics, ranging from nuclear weapons to K-Pop. They used 200 prompt formulations based on research into real chatbot use and generated 70 million responses. The study was conducted by the university’s Department of Computer Science (DIKU).

First author Dustin Wright said people are often exposed to the same information repeatedly, affecting both how they find knowledge and which knowledge they can access. Senior author Isabelle Augenstein, a DIKU professor, said a narrower range of perspectives could allow popular content to become more dominant while other material is overlooked. She described this as a risk of a “vicious cycle.”

Augenstein compared the potential reduction in diversity with globalization, which has made the same products and coffee chains common in many places. She said a knowledge collapse is not happening yet, and that newer models produce slightly more diverse answers than older ones.