Humanizer

A humanizer takes text that a model wrote and rewrites it so detectors are less likely to flag it. Typically it varies sentence length, swaps predictable word choices, adds small irregularities, and generally roughens the smooth, evenly-paced rhythm that gives machine writing away.

It is an arms race with an obvious shape. Detectors learn the signatures of humanizers, humanizers learn what detectors look for, and both sides publish accuracy claims about the other. Worth knowing if you ever get accused of using one: the existence of these tools is part of why detector scores should never be treated as proof.