Heavy reliance on AI chatbots could expose people to a narrower range of information and perspectives, potentially increasing the risk of what researchers describe as “knowledge collapse,” according to a new study led by the University of Copenhagen.
Researchers from the University of Copenhagen and several other institutions tested 27 large language models across 155 topics, using 200 different prompt formulations for each topic. The research generated around 1.7 million AI-generated answers containing approximately 70 million individual claims.
The study found that all the tested AI models produced information that was less diverse than a basic Google search. GPT-5, which generated the most diverse responses among the models tested, still provided at least 18.7% less varied information than Google.
According to the researchers, people who repeatedly rely on chatbots may encounter the same information and perspectives more often, potentially reducing exposure to less prominent or alternative sources of knowledge.
The researchers explain that this lower diversity is partly related to how large language models work. Models compress vast amounts of training data and learn recurring patterns, which can result in less frequently represented information receiving less attention.
The researchers also warned that the effect could become more significant if future AI models are increasingly trained on AI-generated material. However, they stressed that “knowledge collapse” is not currently happening, noting that newer models have generally shown somewhat greater diversity than older ones.
They recommend using AI chatbots as one information source while consulting different sources to gain broader perspectives and a better understanding of complex subjects.