Students are learning to write for AI detectors, not for humans

Facepalm: AI writing detectors were supposed to identify machine-generated text; instead, they are quietly reshaping how students write and how they use AI in the first place. Across classrooms and campuses, tools built on opaque language models and probabilistic pattern matching are pushing some of the strongest writers to tone down their style, study the detectors themselves, and even adopt generative AI defensively just to avoid being flagged.

In one case, an AI checker pre-installed on a school-issued Chromebook flagged a student's essay on Harrison Bergeron by Kurt Vonnegut as "18% AI-written" simply because it contained the word "devoid."

When the student replaced "devoid" with "without," the score dropped to zero, even though the underlying ideas and structure remained unchanged. That behavior is typical of current detection systems, which rely on statistical signals such as word choice and distribution rather than any meaningful understanding of authorship.

As a result ...