How to make AI text undetectable: what is actually possible

Three reasons, each from a primary source.

Detectors are many and they disagree. Weber-Wulff and colleagues ran 54 documents through 12 public tools plus Turnitin and PlagiarismCheck and found the probability of a false positive ranged from 0 percent for Turnitin to 50 percent for GPTZero on the same human-written texts (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). A text that passes one detector has told you nothing about the next.

Detectors move. OpenAI listed, as a limitation of its own classifier in January 2023, that "AI-written text can be edited to evade the classifier" and that classifiers "can be updated and retrained based on successful attacks", adding that "it is unclear whether detection has an advantage in the long-term" (OpenAI, new AI classifier, read 9 November 2026). Three years later, Turnitin states its indicator covers text "modified by AI paraphraser or bypasser (also called humanizers) tools" (Turnitin, AI writing detection FAQ, read 9 November 2026). Whatever worked against a 2023 model is training data for a 2026 one.

The limit runs both ways. Sadasivan and colleagues derived a bound tying the best possible detector's performance to how far apart human and AI text distributions sit (Sadasivan et al. 2023, arXiv 2303.11156, read 9 November 2026). As models write more like people, detection gets harder; as paraphrasers rewrite more, text drifts from what the writer meant. Neither side reaches zero.

A detector does not know who wrote a text. It scores properties of the text and reports a probability. Three properties account for most of what the published detectors describe.

The guide on what AI detectors measure goes through each in depth. The short version: anything that makes word choice less predictable, structure less even and content more specific moves a reading. Those are also the things a careful human editor does to any draft, which is why the honest version of this topic is a guide to humanizing AI text by editing it well.

Two 2023 studies did the measuring that every humanizer vendor now cites, usually without the caveats.

Krishna and colleagues also proposed the defence: an API provider stores what it generates and matches paraphrases back to the source, which in their test caught 80 to 97 percent of paraphrased generations at a 1 percent false positive rate (Krishna et al. 2023, arXiv 2303.13408, read 9 November 2026). That is one more reason the word "undetectable" does not hold: the detection method can change shape entirely.

Nobody in the 2023 literature had an editor read paraphrased output against its input and count what was lost. "Only slightly" and "not appreciably" are model-side judgments. The damage, when it happens, takes three forms: a dropped qualifier that widens a claim, a swapped term that changes it, and a merged sentence that no longer parses. Stronger modes do more of all three.

This is why the published method on this site gives meaning kept 35 points and errors added 20, more than the 30 for detector pass rate. A rewrite that reads as human and says something else has failed, whatever the detector shows.

No detector score should be the only evidence in a misconduct case. The same standard applies to anyone trying to lower one: a reading is not a fact about authorship, in either direction.

Four things, in order. Read the AI policy that applies to you. Where AI use is allowed, disclose it in the form your style guide asks for. Edit the text yourself for facts, structure and word choice, which changes the same signals a tool changes and keeps your meaning. Then, if you want a tool for a second pass, choose one on the fields in the guide to whether AI humanizers work: meaning kept, errors added, detector readings on a stated date, and plan limits.

The guides hub holds the rest of this series. None of it will make text undetectable, because nothing does; all of it will tell you what a given change does to a given detector on a given date, which is the most any honest page can offer.