Chatbots Deliver Less Diverse Knowledge Than Web Search, Study Finds

A study of 27 large language models found that AI chatbots consistently provided users with a narrower range of information than conventional web search. Across 155 topics, every model tested produced less diverse answers than Google, with even the most diverse system, OpenAI’s GPT-5, providing at least 18.7% less varied information. The results do not show that a collapse in knowledge diversity is already underway, but they identify a possible mechanism through which chatbot use and AI-generated training data could contribute to narrower information over time.

The analysis was led by researchers from the Department of Computer Science at the University of Copenhagen, with collaborators at Aalborg University, Stanford University, the University of Colorado Boulder and the University of Texas at Austin. It examined models developed by OpenAI, Meta, Google and Alibaba. The research, titled “What and Whose Knowledge?

Measuring Epistemic Diversity in Large Language Models,” was posted on arXiv and accepted for the international research conference EMNLP 2026, scheduled to take place in October 2026. To test the systems at scale, the researchers created 200 differently worded prompts for each of 155 topics. The formulations were based on questions from real users.

Running these prompts through the 27 models generated a dataset of about 1.7 million AI-generated answers containing approximately 70 million individual claims. The subjects included nuclear weapons, marriage, pornography, racism and genocide, as well as country-specific topics such as Marine Le Pen, the Falklands War and K-pop. The study covered topics relating to 12 countries.

The researchers found that every language model tested supplied more uniform information than a Google search. Across the topics examined, people asking the same broad question in many different ways were repeatedly exposed to substantially the same material rather than the broader range of information obtained through conventional web search. Dustin Wright, the study’s first author and a former postdoctoral researcher at the University of Copenhagen who is now an assistant professor at Aalborg University, said the results show that chatbots affect not only how people find information but also which information they encounter.

“Every language model we tested provides users with more uniform information than a simple Google search across all the topics we looked at,” Wright said. “In other words, people are to a large extent exposed to the same information over and over again. So AI chatbots are not just changing how we find knowledge, but also which knowledge we have access to.”

The researchers linked the observed uniformity to the way large language models operate. Such systems are trained on very large quantities of text and learn patterns in how words and sentences relate to one another. When a user asks a question, the underlying model generates a response by predicting which words are most likely to follow.

According to the research team, this process favors frequently occurring patterns, while information that deviates from those common patterns can be filtered out. Model size did not remove this effect. The analysis found that smaller AI models generated more diverse content than larger models, while newer models were more diverse than older ones.

These results mean that system scale alone did not guarantee a broad range of information, and the tendency toward uniformity was not identical across all models or development periods. Even so, none of the newer or smaller models tested matched the diversity of conventional web search. The finding was based on the researchers’ comparison of model responses with Google results across the selected topics and prompts.

One possible longer-term mechanism is synthetic training data. The researchers expect language models to be trained increasingly on text produced by other AI systems. In that case, models could learn from outputs that are already less diverse than the human-written text on which earlier systems were trained.

If the process were repeated over several generations, the range of information could gradually narrow, a potential outcome the researchers call “knowledge collapse.” The study documented the lower diversity of current model outputs and described this mechanism as a longer-term concern; it did not directly demonstrate collapse by training and testing successive generations of models on AI-generated text. Augenstein said the narrowing of information could create a feedback loop in which popular content becomes even more dominant while other content is increasingly overlooked. She compared the process with globalisation, noting that the worldwide availability of the same products and coffee chains has advantages but has also reduced diversity.

This is a proposed consequence rather than a measured result of the study, which did not quantify the size of such a feedback loop. “It’s a worrying thought. However, we can see that the more recent models produce slightly more diverse answers than older models, so knowledge collapse is not happening yet,” said Isabelle Augenstein, senior author and a professor at the University of Copenhagen.

“But the mechanism that could trigger it in the longer term is already there. So it is something we should be aware of.” Rather than dismissing chatbot summaries as unhelpful, Augenstein emphasized that they can be useful while remaining open to wider information.

“AI chatbot summaries can of course be useful – that is why so many people use them. But it is still important to seek out different sources in order to understand the nuances and get a broader picture – especially for the younger generation growing up with AI,” she said. She also warned against becoming so dependent on the technology that people stop understanding and thinking for themselves.

The research compared the diversity of information delivered by chatbots and web search; it did not test individual learning outcomes or establish how frequently people use each service as their primary source of knowledge. The researchers also called on AI developers to preserve the breadth of knowledge represented in their systems. Wright said the team has developed a method that developers can use to measure diversity in language models, which could help ensure that future systems do not become less diverse.

The immediate finding is therefore both technical and practical: the 27 chatbot systems produced measurably more uniform information than Google across every topic tested, although newer models showed slightly greater diversity than older ones and smaller models were more diverse than larger ones.

Subject of Research: Epistemic diversity in large language models compared with conventional web search.

Article Title: AI chatbots give us a narrow slice of knowledge: Researchers warn of ‘knowledge collapse’

Article References: AI chatbots give us a narrow slice of knowledge: Researchers warn of ‘knowledge collapse’. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: AI chatbots, large language models, epistemic diversity, information diversity, web search, knowledge collapse, synthetic training data, OpenAI, Google, language model research, chatbots, give