Bot detectors mistake 39.1% of agents for humans

Binary bot detection mislabeled 39.1% of real browser agents as human with one classifier and 34.5% with another in a controlled July 29 study of AI-agent traffic. Adding an explicit agent class produced an agent F1 of 1.000 across 30 runs, although the signal came from Playwright’s missing physical-input events—not proof that the system recognized reasoning—and may weaken if automation stacks change. As native computer-use models move browser and desktop actions into production workflows, operators should give sanctioned agents an authenticated identity instead of relying on behavioral bot filters to infer one.