AI chatbots may narrow what people know, Danish study finds

For millions of people, asking an AI chatbot has become as routine as typing a query into Google. Whether it is finding an explanation for a difficult concept, drafting an email, understanding an economic issue or deciding what to cook, chatbots increasingly serve as the first stop for information.

But a new study led by researchers at the University of Copenhagen suggests that this convenience may come with a hidden cost: users could be exposed to a much narrower range of information and viewpoints.

Researchers at the university’s Department of Computer Science (DIKU) tested 27 large language models across 155 topics connected to 12 countries. The study generated around 1.7 million chatbot responses containing about 70 million individual claims.

The researchers found that every chatbot tested produced more uniform information than a conventional Google search across the topics examined. In practical terms, users asking different chatbots the same or similar questions were more likely to encounter repeated claims, information and perspectives.

Even the most diverse model in the study, OpenAI's GPT-5, provided at least 18.7% less varied information than Google, according to the researchers.

Chatbots offer answers, but fewer perspectives

The study's concern is not that chatbot responses are necessarily shorter. Instead, researchers focused on the variety of information users encounter.

A traditional search engine can direct users to numerous websites, publications and sources, potentially exposing them to different interpretations, evidence and viewpoints. Chatbots, by contrast, typically generate a single consolidated response, which can make similar claims and angles appear repeatedly.

"Every language model we tested provides users with more uniform information than a simple Google search across all the topics we looked at," said first author Dustin Wright, a former postdoctoral researcher at DIKU who is now an assistant professor at Aalborg University.

The result, he said, is that AI is changing not only how people search for knowledge but also the range of knowledge they encounter.

Smaller models sometimes showed greater variety

The researchers also found differences between AI models. Smaller models sometimes generated more varied responses than larger ones, while newer models generally performed better than older systems.

However, none of the models matched the diversity offered by a search engine that directs users to multiple independent websites.

The researchers argue that this matters because repeated exposure to the same information can gradually reduce the range of perspectives people encounter, particularly when AI becomes their default source for learning and decision-making.

"We risk exposing people to fewer perspectives and a narrower range of knowledge," said Isabelle Augenstein, a professor at DIKU and the study's senior author.

She warned that this could create a cycle in which already popular information becomes increasingly dominant while less prominent material receives less attention.

What ‘knowledge collapse’ means

The researchers use the term "knowledge collapse" to describe a potential long-term consequence of this narrowing.

If future AI models are increasingly trained on material generated by other AI systems, the researchers argue, the lack of diversity could compound from one generation of models to the next.

In such a scenario, AI-generated information could become progressively more similar, reducing the breadth of knowledge available to users.

The study does not suggest that AI chatbots should be abandoned. Instead, researchers emphasise the importance of using them alongside other sources of information.

What students and users can do

For students in particular, researchers recommend checking chatbot-generated information against official websites, textbooks and multiple reliable sources, especially when dealing with examination syllabi, government schemes, admission rules and other factual requirements.

Users can also use search engines to explore different websites and perspectives rather than relying on a single AI-generated answer.

Classroom teaching, discussions with peers and library resources can further expose learners to ideas that may not appear in a chatbot's response.

AI chatbots remain powerful tools for drafting, summarising and quick revision. But the Copenhagen study suggests that convenience should not come at the expense of curiosity and variety.

The central concern is simple: if people repeatedly ask machines for answers and receive increasingly similar information, they may gradually encounter less of the wider world of knowledge that exists beyond those answers.