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

The 90-day recency window is not just a design choice, it is the entire argument: by restricting queries to events that postdate most training cutoffs, LiveBrowseComp removes the escape hatch that lets models fake retrieval by surfacing memorized answers. The benchmark is essentially a lie detector for search agents.

This story is largely disconnected from recent activity in our archive, as we have no prior coverage of search-agent evaluation methodology or the Harbin Institute team. It does, however, belong to a broader and increasingly active conversation about the gap between benchmark performance and real-world utility in deployed AI products. The core tension here, that models optimized on static corpora may be structurally ill-suited for live information tasks, has implications for every product currently marketed as an AI research assistant. The reshuffling of capability rankings is worth taking seriously because it suggests prior leaderboard results on web-search tasks may have been measuring memory, not reasoning.

Watch whether OpenAI or the Kimi team publish a direct response to LiveBrowseComp scores within the next 60 days, either by disputing the methodology or by releasing updated agents with verifiable retrieval traces. Silence from both would itself be informative.