The campus divide over AI detectors and proof of cheating

A UW–Madison biology professor told The Atlantic he planned to scrap writing assignments: students had learned to remove his hidden chatbot-catching prompts, and university guidance discouraged relying on detectors alone.

The magazine cites a University of Chicago working paper that found Pangram essentially never mistook long human-written passages for AI, though it occasionally errs on shorter snippets. It also describes a Notre Dame study that found “humanizer” tools could fool an older version; Pangram says its latest model is more robust against them.

Separately, The Dartmouth reports that nine of Provost Santiago Schnell’s works published in 2026 received a median “AI-written” text percentage of 96% from Pangram. Schnell told the paper he uses AI for language refinement and editing, but reviews the final text and takes responsibility for its content. He said a detector’s classification is not a determination of authorship.