In The Detection Inversion, I argued that better RLHF training makes safety harder to verify. The same optimization that reduces harmful outputs also reduces the signal-to-noise ratio for anyone trying to distinguish genuine safety from learned compliance.
Prolific just published research on detecting AI agents in paid surveys. Their mouse tracking catches 95% of agents. Classic attention checks? 59%. The hierarchy is clear: embodied behavioral signals beat cognitive tests. Agents averaged 6-10 mouse movements per page; humans made 85. Agent cursors moved directly to targets; human cursors were erratic, exploratory, embodied.