AI writing detectors are actively destroying careers, and the tools doing the damage admit they don’t work. A 2023 Stanford study confirmed that detectors like GPTZero and Turnitin’s built-in scanner disproportionately flag non-native English speakers as AI authors, a bias that maps directly onto a global workforce. Forty-three percent of US teachers already use these tools regularly, publishers are canceling seven-figure book deals<\/a> on their basis, and students are winning lawsuits against universities that acted on flagged results. MIT and Yale have formally warned that the tools don’t work.<\/p>\n Your exposure here isn’t abstract. Any organization using AI detection software to audit employee communications, written deliverables, or hiring materials is operating a liability engine with a known defect rate. The vendors themselves disclaim reliability, Turnitin explicitly says its tool “may not always be accurate” and shouldn’t drive action against a student, yet 43 percent of educators are using it as exactly that. Translated to an HR context: a performance review process that flags “AI-assisted writing” is one wrongful termination suit away from a very public lesson in probabilistic inference.<\/p>\n The bias pattern compounds the exposure. Non-native English speakers and neurodivergent writers are systematically more likely to trigger false positives, the statistical term for a detector calling human writing AI when it isn’t. That demographic overlap with diversity hiring targets isn’t coincidental; it’s a structural problem baked into how detectors are trained. An enterprise that disciplines employees based on detector output, without independent corroboration, is essentially using a biased instrument to evaluate a protected class. The detector companies’ rebuttals to the Stanford study are exactly what you’d expect from vendors whose entire business model requires the tools to appear credible.<\/p>\n The smarter bet is what MIT and Chicago are already modeling: shift the frame from detection to disclosure. Design workflows and performance standards around whether AI was used appropriately for a given task, not whether a probabilistic scanner says it might have been. CHROs who build that policy architecture now will be insulated when the first major enterprise discrimination case involving AI detection tools lands, and that case is coming. The only falsification condition that would change this view is a detector that publishes audited, third-party false-positive rates stratified by writer demographics. None currently does.<\/p>\nWhat this means for your business<\/h2>\n