Kenneth Glad
Reaching for your phone to ask an AI chatbot a quick question has become second nature for millions of people. It feels like having the entire world’s knowledge right at your fingertips, packaged in a neat summary. Yet beneath those convenient answers lies a quiet narrowing of what we actually learn.
Turning to AI for everyday answers is fast replacing the old habit of scrolling through search results. Whether planning dinner, drafting a tough email, or trying to make sense of interest rates, people increasingly rely on chatbots to do the thinking.
Recent research shows that this shift comes with a hidden downside. Large language models provide a far narrower range of information than a standard web search.
According to a study led by researchers at the University of Copenhagen, every single language model tested returned more uniform information than a basic Google search across all subjects.
“In other words, people are to a large extent exposed to the same information over and over again,” explained Dustin Wright, an assistant professor at Aalborg University and lead author of the study. He noted that AI is not just changing how we find facts, but also which facts we get to see.
Narrowing the web
The team tested 27 different language models across 155 subjects, running 200 distinct questions per topic to build a dataset of around 70 million claims. Topics ranged from nuclear weapons and racism to regional history and pop culture.
Even the most varied AI model tested, OpenAI’s GPT-5, proved to be at least 18.7 percent less diverse in its responses than Google search.
Researchers warn that relying solely on AI could lock society into an echo chamber. The most popular perspectives dominate, while lesser known views vanish.
Senior author Isabelle Augenstein compared the trend to modern global retail chains, where the exact same options appear everywhere. “That has many advantages, but it has also reduced diversity,” she noted.
AI models create this uniformity because they compress vast text datasets, learning to repeat the most frequent patterns while filtering out rare details. As future models train on text written by earlier AI, the pool of unique information could shrink even further, creating what researchers call a knowledge collapse.
While newer models show slight improvements, experts stress that users must continue cross-checking multiple sources to keep a broad perspective.
Source: University of Copenhagen
