Angel Ortmann Lee examines cognitive surrender and automation bias in human-in-the-loop systems. In a Duolingo English Test experiment using legitimate historical sessions, experienced proctors accepted about half of fabricated copy-typing alerts. The speaker emphasizes that the experiment did not affect test takers.
Angel Ortmann Lee describes a guideline change telling proctors that an AI flag was preliminary and that independent video evidence was required before upholding it. She reports that rejection of the false alerts rose from about 50% to 71%, illustrating how interaction design can matter without changing the detection model.
Angel Ortmann Lee separates whether an algorithm correctly detects an object from whether that object constitutes a policy violation. A hearing aid, for example, should not be treated as cheating simply because a detector labels it as an earbud. Combining those judgments into one button also contaminates training feedback.
Angel Ortmann Lee extends the argument to writing tutors and coding agents: feedback should be targeted, assumptions visible and work divided into meaningful reviewable steps. High-stakes decisions need deliberate friction, while low-stakes interactions can be smoother. Recording overrides and explanatory feedback produces richer evaluation data than a stream of accepted or rejected suggestions.
Watch on YouTube
