AI Detectors
An AI detector estimates whether a text was generated by AI — the function AI humanizers exist to defeat. This site catalogues the humanizer side of that arms race; this page records what the credible evidence says about the detector side.
Why a humanizer site needs this page
Every humanizer claim is a claim about detectors: "undetectable" means "detectors miss it". So the detector evidence base is not optional context — it is the ceiling on what any humanizer claim could even mean. And that evidence base is more solid on the detector side than anything published on the humanizer side: two peer-reviewed or working-paper studies measure detector behavior on humanized text directly, which is more independent evaluation than any humanizer in this market has received.
What the studies found — attributed, in full
Van Vlasselaer, Van Droogenbroeck and Spruyt (International Journal for Educational Integrity, June 2026): four detectors (GPTZero, Copyleaks, Turnitin, Pangram) on 160 academic papers of at least 4,000 words — human-written, AI-generated, mixed, and humanized. Accuracy on humanized text: Pangram 92.5%, Turnitin 50.0%, Copyleaks 22.5%, GPTZero 2.5%. Jabarian and Imas (NBER working paper 34223, September 2025): Pangram, Originality.ai, GPTZero, and an open RoBERTa detector across genres, lengths, and models — Pangram's error rate stayed near zero. Note what these numbers do not tell you: nothing about which humanizer is "best" — the same point our methodology page makes about why this site refuses to rank humanizers on bypass figures.
The arms race, stated plainly
A humanizer rewrites text to move it away from the statistical patterns detectors measure; a detector vendor then retrains against rewrites; and so on. Two honest conclusions fall out of the record. First, any fixed "bypass rate" is a snapshot of a moving target measured by someone with a stake — which is why we exclude such figures from scoring. Second, the spread in the detector studies (92.5% versus 2.5%) means "AI detection" is not one thing — detector choice changes outcomes more than humanizer choice does, by the only credible numbers on file.
FAQ
Is an AI detector the opposite of an AI humanizer?
Yes — they are the two sides of one arms race. A detector estimates whether text was generated by AI. A humanizer rewrites text so detectors stop flagging it. This site catalogues the rewriting side and scores vendor transparency; detector-side tools are evaluated by related detector publications.
Which AI detectors work well on humanized text?
The peer-reviewed record shows enormous spread: a June 2026 study in the International Journal for Educational Integrity measured four detectors' accuracy on humanized academic text at 92.5% (Pangram), 50.0% (Turnitin), 22.5% (Copyleaks), and 2.5% (GPTZero), and an NBER working paper (34223) found Pangram's error rate near zero across genres. Those figures describe detectors, not humanizers — and we cite them with attribution rather than as our own measurements. For detector rankings, see the related publications in our footer.
The other senses: our AI humanizer catalogue and our paraphrasing tool page. Detector-side evaluations: see related publications in the footer.