UF researchers tested the five most popular AI text detectors and found false negative rates as high as 99.6%. A single vocabulary tweak defeated most of them entirely.