AI search agents often confirm what they already know instead of actually researching the web

Researchers from Harbin Institute of Technology and Xiaohongshu tested 11 AI models on search tasks and found many performed well even without web tools, suggesting heavy reliance on internal knowledge.

When search was enabled but relevant documents were removed, performance often fell below closed-book results, showing that search can sometimes distract models from the right answer.

To measure real browsing ability, the team introduced LiveBrowseComp, a time-sensitive benchmark built from recent facts unlikely to be in training data.

The study argues that current AI search benchmarks may overstate genuine web research skill by rewarding memorized knowledge instead of real information gathering.