Do AI humanizers actually work? A study of 1,485 real rewrites

We measured the GPTZero score of every rewrite our users ran over one week. Here is what the data shows — and what it does not.

Updated October 17, 2026

Across 1,485 real rewrites logged on humanize.beer between 10–17 October 2026, 87% of humanized outputs landed in GPTZero’s likely-human band (under 30% AI), with a median AI probability of 0.2%. Results were strongest in the two-pass Academic mode (92% under 30%). This measures GPTZero only — not Turnitin or a human grader — and no score guarantees any other detector’s verdict.

This is our own data, from our own tool. We publish it because honest numbers are more useful than marketing claims — including the places where humanizing works less well.

What we measured

Every time humanize.beer rewrites text, it immediately scores the result with GPTZero and shows you an AI-vs-human probability. With consent, that result is saved to an anonymised dataset (see our privacy page; you can opt out). We analysed every saved rewrite from 10–17 October 2026 that carried a GPTZero score.

• 1,661 rewrites were logged in that week; 1,485 had a GPTZero score and form this analysis.
• These are real user texts, not a curated test set — mostly AI-assisted drafts people wanted to read more naturally.
• Median input length was 196 words.
• The metric is GPTZero's AI-probability for the humanized output. Lower means more human-like to GPTZero.

The headline result

Across all 1,485 scored rewrites, the median GPTZero AI-probability was 0.2%, the mean was 12% (a few high-scoring outputs pull the average up), and 87% of outputs scored under 30% AI — the band GPTZero treats as likely human.

Results by rewrite mode

The Academic and GPTZero modes both land the large majority of rewrites in GPTZero's likely-human band. ZeroGPT mode is the honest outlier: it is tuned to lower scores on ZeroGPT, a different detector, so its GPTZero numbers are naturally higher. The lesson is practical — pick the mode for the detector you actually care about, and don't assume one detector's score transfers to another.

What this does and doesn't prove

• It measures one detector. GPTZero is widely used, but it is not Turnitin, Originality.ai, Copyleaks or a human reader. A low GPTZero score does not predict any of those.
• It is an output snapshot, not a controlled before/after. We report where humanized text lands, not a guaranteed drop for any single document.
• Detectors are probabilistic and change. GPTZero updates its model; false positives and false negatives are real, especially on short or non-native-English text. A score is a signal, not proof.
• Reading well matters more than any score. The point of humanizing is text that genuinely reads like a person wrote it — which is also what survives a detector update.

Try it on your own text

The most useful test is your own paragraph. Paste a draft into the humanizer, read the result aloud, and check the AI detector score. For how these detectors judge text in the first place, see how AI detectors work and how GPTZero, Turnitin and Originality differ.

Frequently asked questions

Keep exploring

• How AI detectors workPerplexity, burstiness and classifiers, explained simply.
• Best free AI humanizersWhat is actually free, and which tool fits which job.
• GPTZero vs Turnitin vs Originality.aiWho uses each detector and how they report scores.