Universities increasingly turn away from AI detection tools

Universities are pulling back from AI detection tools even as some products claim gains in accuracy, leaving instructors to reconsider writing assignments and assessment methods. Will Oremus reports for The Atlantic that concerns about false accusations, student privacy, and an escalating battle with evasion tools are limiting their adoption.

Timothy Paustian, a biology professor at the University of Wisconsin at Madison, says he identified 60 AI-written submissions in one online microbiology assignment with 350 students. He used both detection software and hidden instructions intended to trigger chatbot responses. Yet his university advises faculty not to rely on detectors alone, and students have learned to remove hidden prompts. Paustian now plans to eliminate writing assignments from that course.

Accuracy does not settle the dispute

Turnitin, the most widely used provider in higher education, says its system has a false-positive rate below 1 percent and misses about 15 percent of AI-generated writing. The company says its results should begin a conversation, not serve as conclusive evidence of misconduct.

That distinction matters at scale. Vanderbilt University previously estimated that screening 75,000 papers annually could still wrongly flag up to 750 papers at a 1 percent false-positive rate. Some institutions, including Vanderbilt and Indiana University, have restricted or barred detector use.

Pangram, a newer tool, claims substantially lower false-positive rates. A University of Chicago working paper found it rarely identified longer human-written passages as AI. However, researchers at Notre Dame report that “humanizer” tools can alter AI-generated text to evade detectors. Students can also test and revise text against publicly available systems such as Pangram.

Meanwhile, educators face a broader policy choice. MIT recommends reducing reliance on detection and redesigning assignments around process, reasoning, and work with clear value beyond a grade. But such changes require time and staffing that many instructors, particularly those managing large online classes, do not have.