Why "undetectable" is a marketing word, not a technical one

"Undetectable" cannot be true of any text, because AI detectors are probabilistic classifiers that are retrained and re-thresholded without notice. A passage that scores 4% AI today can score 40% next month on the same tool, and differently again on a different one. Any product that promises undetectability is describing a snapshot as if it were a law.

What a detector actually does

A detector estimates how predictable a text is to a language model and how much that predictability varies. It outputs a probability. The vendor then chooses a threshold (the score above which the tool says "likely AI") and moves that threshold whenever they want to trade false positives for false negatives. Turnitin has done this publicly. GPTZero has changed models several times. The number you see is the product of a model, a threshold and a date.

So "passes GPTZero" is a statement about one run on one day. It is not a property of the text.

What the word hides

The marketing word hides three things. First, that the result was measured once, usually by the vendor selling it. Second, that detectors disagree with each other by wide margins, so passing one says little about another. Third, that the tools are adversarial: every rewriting technique that works at scale becomes training data for the next detector version.

It also hides a quieter cost. Products that sell undetectability attract users whose goal is to evade a rule rather than to write well, and they get sued, delisted, or cut off by payment providers. That is not a business we want to be in.

What we say instead

AntiGPT rewrites AI-assisted text so it reads naturally in your voice, and shows you a detector's estimate before and after so you can see the effect of your edits. We publish measured results only when we have measured them, with the date and the method, and we put a fixed disclaimer next to every score: it is an estimate, it changes, and no text can be guaranteed to pass anything. If a tool tells you otherwise, ask when they measured it, and on which version.