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

In three linesAI search agents like GPT-5.4 and Kimi K2.6 mostly confirm their training knowledge rather than genuinely researching the web. Researchers at Harbin Institute of Technology demonstrated this using LiveBrowseComp, a benchmark based on events from the last 90 days. Without relying on training memory, performance collapses.

Summary generated by Claude — human-verified