Learning With AI Falls Short Compared To Old-Fashioned Web Search

November 24, 2025

Researchers from the University of Pennsylvania conducted seven studies with over 10,000 participants to examine how learning through large language models like ChatGPT compares to traditional Google searches. The experiments consistently showed that people who used AI chatbots to learn about topics developed more superficial understanding and produced shorter, less useful advice than those who navigated web links themselves. This difference persisted even when researchers controlled for the information presented and the platform used, suggesting the problem stems from LLMs transforming learning from an active to passive process.

Who is affected

Students in secondary education facing challenges in developing foundational skills
Millions of people using large language models to access knowledge since late 2022
More than 10,000 research participants in the studies
Independent readers who evaluated advice written by study participants
Educators trying to balance skill development with AI integration
People seeking to develop deep, generalizable knowledge in specific areas

What action is being taken

Researchers are studying generative AI tools that impose "healthy frictions" for learning tasks
People are using large language models like ChatGPT to access knowledge
Researchers are examining which types of guardrails or speed bumps successfully motivate users to actively learn beyond synthesized answers

Why it matters

This research reveals a significant trade-off in how people acquire knowledge using AI tools. When learners rely on pre-synthesized LLM responses, they bypass the cognitive friction of navigating, interpreting, and synthesizing information themselves—a process fundamental to developing deeper understanding. The passive nature of LLM-based learning results in knowledge that is shallower, less original, and less useful to others. This has important implications for education, professional development, and decision-making, as it suggests that the convenience of AI may undermine the quality of learning and expertise development in society.

What's next

The researcher plans to study generative AI tools that impose healthy frictions for learning tasks
Future research will examine which specific types of guardrails or speed bumps most successfully motivate users to actively learn beyond easy, synthesized answers
Development of such tools is identified as particularly critical for secondary education

Read full article from source: The San Diego Voice & Viewpoint