Universities grapple with AI-generated homework, abandoning unreliable detection tools

At the University of Wisconsin-Madison, biology instructor Timothy Paustian tried embedding a secret cue in an assignment to expose AI-written work, yet only a fraction of students were caught and newer models bypassed the trick. He now argues that traditional writing tasks are becoming ineffective for lower-level courses, urging more creative assignments. Across the country, professors at Yale, Cornell and other schools are abandoning AI detection software because of inaccuracy, bias against non-native speakers, and the rise of tools that rewrite AI text to appear human.

The unreliability of detectors has led institutions to pair written work with oral examinations or video debates, though concerns remain that students could still use AI to generate content. Researchers highlight a growing technology arms race between detection systems and “humanising” programs, while a recent MIT report warns that easy chatbot answers may foster “cognitive surrender” among learners. Some academics, like Harvard College dean David Deming, even advocate embracing AI in writing courses rather than fighting it.