AI search agents confirm existing knowledge rather than truly researching the web
Leading AI search agents (GPT-5.4, Kimi K2.6) primarily confirm training data via web searches rather than performing genuine research, finds Harbin Institute of Technology
New time-based benchmark (LiveBrowseComp) testing only 90-day-old events exposes severe performance degradation when models can't rely on training memory
Current AI agent rankings questioned: actual research capabilities collapse without access to pre-trained knowledge