SIGNAL
Universities did not ask for an AI misconduct signal that punishes the student who disclosed their AI use and rewards the one who ran their draft through a humanizer. They got one anyway. Detector-led integrity policy treats a stand-in for pre-LLM prose as ground truth, and that stand-in is exactly what an honest AI edit produces. Lower the score, the threshold moves; raise it, the humanizer-assisted rewrite walks through.
The arXiv study 'Why AI Detection Fails for Academic Integrity' tested this cleanly. Light "refine abstract only" edits, a proxy for guideline-compliant AI use, were flagged at 64 to 80 percent by Pangram and GPTZero. Unmodified 2023–2025 originals sat at 9 to 15 percent. After Undetectable AI humanization, fewer than 4 percent of AI-labeled rewrites remained flagged. The honest editor carries roughly 16 to 20 times the sanction risk of the humanizer-assisted cheat. The detectors were not measuring authorship. They were measuring proximity to a 2014 academic abstract.
This is a category error dressed as a misconduct tool. Detector scores trained against a pre-LLM baseline will keep ranking students by proximity to that baseline, not by whether they cheated. Universities running integrity policy off a single detector score have built a tax on rule-followers and a subsidy for cheaters who know which browser tab to open. The lever is not better detection. It is dropping the score as standalone evidence and replacing it with process-level evidence the detector cannot fake. Until then, the institution should be defending the student who followed the rules, not flagging them.