“Let me just ask the chatbot.” For many of us, it has become an automatic reflex, whether we are planning dinner, wording a tricky email or trying to understand what rising interest rates mean. But a new study led by the University of Copenhagen suggests that this convenience comes with a hidden cost: the answers we get are much more alike than the ones a simple web search would give us.
What the researchers did
A team from the Department of Computer Science (DIKU) at the University of Copenhagen, together with colleagues from Aalborg University, Stanford, the University of Colorado Boulder and the University of Texas at Austin, tested 27 large language models from OpenAI, Meta, Google and Alibaba. They covered 155 topics relating to 12 countries, from nuclear weapons, marriage and racism to Marine Le Pen, the Falklands War and K-pop.
For each topic, the models were given 200 different phrasings of questions based on real user queries. Together, this produced around 1.7 million answers containing roughly 70 million individual claims. The researchers then measured how varied that information was and compared it with the diversity of a conventional Google search.
The main finding
Every model tested gave users more uniform information than a search engine, across every topic. Even the best performer, OpenAI’s GPT-5, offered information that was at least 18.7% less varied than Google. Two smaller patterns also emerged: smaller models tended to produce more diverse content than larger ones, and newer models were somewhat more diverse than older ones.
“So AI chatbots are not just changing how we find knowledge, but also which knowledge we have access to,” says first author Dustin Wright, now an assistant professor at Aalborg University.
Why are the answers so similar?
The reason lies partly in how language models work. They compress enormous amounts of text and learn the patterns that appear most often. Information that departs from those common patterns tends to be filtered out along the way. The result is answers that gravitate toward the mainstream view.
The researchers compare it to globalisation: the same products and the same coffee chains now appear almost everywhere. That brings real advantages, but it also reduces diversity.
What is ‘knowledge collapse’?
The concern grows when AI models are trained on text written by other AI models, which the researchers expect to become more common. Those models would learn from outputs that are already less diverse than human writing. Repeated over several generations, the range of available information could gradually narrow. This is what the team calls knowledge collapse.
The authors are careful to add that this is not happening yet. Newer models actually show slightly more diversity than older ones. But, as senior author Professor Isabelle Augenstein points out, the mechanism that could trigger it is already in place.
What can we do about it?
The researchers do not argue against using chatbots, since their summaries are clearly useful. Their advice is to use them thoughtfully:
- Consult more than one source to see the nuances and get a broader picture, especially for topics where perspectives differ.
- Keep thinking for yourself. The authors are particularly concerned about younger people growing up with AI becoming too dependent on it.
- Ask developers to protect breadth. The team has created a method for measuring diversity in language models, which AI developers could use to make sure future models do not become less diverse.
Worth keeping in mind
This study compares language models with a Google search, and it measures diversity of claims across the topics chosen by the researchers. It does not say that chatbot answers are wrong, only that they are more uniform. The paper has been accepted at the international EMNLP 2026 conference in October 2026, which reflects peer review by the field, though the results will likely be debated as models keep changing quickly.
Sources
- Original paper: Wright D, Augenstein I, et al. What and Whose Knowledge? Measuring Epistemic Diversity in Large Language Models. arXiv, DOI: 10.48550/arXiv.2510.04226. Accepted at EMNLP 2026. Read the paper
- Press release: University of Copenhagen, AI chatbots give us a narrow slice of knowledge: Researchers warn of ‘knowledge collapse’, September 28, 2026, via EurekAlert!. Read the press release
Leave a Reply