Large Language Models (LLMs) provide information rapidly but diminish critical thinking and understanding among users by stripping away crucial context, according to recent academic research.
This efficiency compresses the traditional information gathering process, offering answers without the “path metadata” — such as friction, elapsed time and source variety — that typically helps individuals assess an answer’s reliability and depth.
Research from Wharton, published in PNAS Nexus in October 2025, found that participants who used AI summaries for learning demonstrated less knowledge, engaged less with the material and produced sparser, less original advice, even when presented with identical factual information.
Shiri Melumad and Jin Ho Yun, both researchers at Wharton, conducted the study.
Separately, the Pew Research Center reported in March 2025 that when AI summaries appeared in Google search results, users clicked fewer normal search results, dropping from 15 percent to 8 percent. Users also clicked cited sources less frequently, at only 1 percent, and ended their browsing sessions more often, with a 26 percent rate compared to 16 percent.
Earlier findings from Yale researchers in 2015 indicated that searching the internet inflated individuals’ belief in their own knowledge. This phenomenon, often termed the “Google effect,” confused access to information with actual understanding, even when searches yielded no results.
A 2025 paper by Microsoft Research and Carnegie Mellon further suggested a correlation between greater confidence in AI-generated information and reduced critical thinking among knowledge workers.
Dirk Lewandowski, an expert in information science, noted that the lack of transparent sourcing in AI-generated responses can obscure the effort and variety of sources that would typically inform a user’s judgment of information quality.
The combined findings suggest a growing concern that while LLMs offer unprecedented access to synthesized information, they may inadvertently foster a shallower engagement with knowledge and an overestimation of personal understanding.
Source: Search Engine Journal
