If you have swapped web searches for an AI chatbot, you may be seeing a smaller slice of the world. A large new study from the University of Copenhagen found that AI language models consistently return less varied information than a plain Google search.
What the researchers did
The team tested 27 large language models from OpenAI, Meta, Google and Alibaba. They asked about 155 topics related to 12 countries, using 200 different phrasings per topic based on real user questions. In total, they analyzed about 1.7 million AI responses containing some 70 million individual claims, then compared how diverse that information was with what web search returned.
What they found
Every model came out less diverse than search. Even the most varied one, OpenAI’s GPT-5, provided at least 18.7% less varied information than Google.
“Every language model we tested provides users with more uniform information than a simple Google search,” said first author Dustin Wright of Aalborg University. “People are to a large extent exposed to the same information over and over again.”
There were some surprises. Smaller models produced more diverse content than larger ones, and newer models were more diverse than older versions. But all of them remained well behind web search.
Why “knowledge collapse” matters
The researchers warn of a possible feedback loop. As more text online is written by AI, future models may be trained on it, and each generation could narrow the range of knowledge a little more. They call this risk “knowledge collapse.”
The catch
The study measures how varied the information is, not whether it is accurate. A narrower answer can still be correct, and a wider spread of search results can include low-quality pages. The comparison also uses Google search as the benchmark, and the findings cover the specific models and topics tested. The work has been accepted at the EMNLP 2026 conference, and a preprint is available on arXiv.
What to do about it
The fix is simple and familiar. “It is still important to seek out different sources in order to understand the nuances and get a broader picture,” said senior author Isabelle Augenstein of the University of Copenhagen. Treat a chatbot’s answer as a starting point, then check other sources — especially on topics where viewpoints differ.