AI makes knowledge superficial, and here's why
A new study has shown that a person will gain more superficial knowledge when using AI to retrieve information on a specific topic than if they did so using a standard Google search.
A new study has shown that a person will gain more superficial knowledge when using AI to retrieve information on a specific topic than if they did so using a standard Google search.
A new study has shown that a person will gain more superficial knowledge when using AI to retrieve information on a specific topic than if they did so using a standard Google search.
These findings are based on an analysis of seven studies involving more than 10,000 participants, Futurism writes .
The essence of the research was as follows: participants were given the task of studying a topic, divided into two groups - those who were to use only an artificial intelligence chatbot for research, and those who used a standard search engine. At the end, they were asked to write a second piece of advice based on the material studied.
A clear pattern emerged: participants who used AI for their research wrote shorter tips that contained general recommendations and less factual information, while humans who used Google search provided more detailed and thoughtful advice.
This pattern held even after accounting for factors such as the information users saw during the study, by showing each group the same facts or tools they used.
“The results confirmed that even when the facts and platform were the same, learning based on LLM synthesized answers resulted in more superficial knowledge compared to collecting, interpreting, and synthesizing information for oneself through standard web links,” said study co-author Shiri Melumad, a professor at the Wharton School of the University of Pennsylvania.
She explains that one of the main principles of skill acquisition is that people learn best when they actively interact with the material. For example, when researching a topic through Google, people overcome much more “resistance”: looking at different links, reading sources, interpreting and summarizing information on their own.
“But with large language models,” added Melouad, “this entire process is done for the user, transforming learning from a more active to a passive process.”



