HumanizerEval, October 2026 review
How HumanizerEval tests AI humanizers
HumanizerEval is a public benchmark for AI humanizers. Its October 2026 review, run on October 22, 2026, put 11 humanizers through the same 50 texts and scored every output with 6 AI detectors. This page explains how that test works, following HumanizerEval's own methodology.
Disclosure: HumanizerEval is built by the StealthGPT team, and StealthGPT is one of the humanizers it ranks. Its per-sample scores are public, so you can check the results yourself at humanizereval.com.
The same texts for every tool
Each review uses one set of English texts written by AI models for the benchmark. In October 2026 that was 50 texts of 364 to 997 words. Every humanizer got the same texts, and every humanized output went to the same detectors.
The source texts are published on HumanizerEval with the prompt and model behind each one. The humanized outputs are not.
Humanizer settings
Each tool runs on its strongest advertised setting on its cheapest paid plan. HumanizerEval records the exact mode, model and plan for every tool in every review, StealthGPT included, and lists them on its methodology page.
The detectors
Every output is scored through each detector's own API. Detector versions are pinned within a review and the model setting is recorded for each run, because detector updates are the main reason rankings move between reviews. These are the 6 detectors in the October 2026 panel:
| Detector | API model setting |
|---|---|
| Pangram | v4 |
| GPTZero | latest |
| Originality.ai Turbo | turbo |
| Originality.ai 5.0 | model-5.0 |
| ZeroGPT | default |
| Winston AI | latest |
What counts as a pass
A humanized text passes a detector, which HumanizerEval calls a strict bypass, when the detector gives it a human score above 50%. The vendor's label doesn't change that: a text marked "mixed" can still pass, and a score of 50% or lower is a fail.
Each detector's rate is the share of the 50 texts that passed. HumanizerEval also publishes mean and median human scores on each detector's page, because a tool can clear the line narrowly on many texts or decisively on fewer.
Human scores are specific to each vendor, so HumanizerEval never averages one detector's score with another's. Originality.ai 5.0 reports the share of a text it judges AI-written rather than a confidence score, so its numbers can't be compared with Originality.ai Turbo's.
How the ranking works
Rank is the mean of the 6 per-detector pass rates. There is no weighted score and there are no penalty points. When two tools tie, the one that passed every detector on more texts ranks higher.
With 50 texts per tool, the 95% confidence interval on a pass rate is roughly 14 points either way, so a gap of a few points between two tools is noise.
Quality is reported separately
Every output is also checked for factual consistency, naturalness, syntax and stance: whether facts and names stay accurate, whether the wording reads fluently, whether sentences are complete, and whether the original position survives. A pass means no issue was found. These results have their own table on HumanizerEval and never feed into the rank.
What isn't tested
Turnitin has no public API at the volume a benchmark needs, so it isn't in the panel. Copyleaks is a candidate for a later review. StealthGPT's own AI detector isn't in the panel either.
Checking the numbers
Every tool's per-sample detector scores are in a public CSV, and a script in HumanizerEval's GitHub repository recomputes each mean and pass count from it. Published reviews are never edited afterwards.
No tool can pay for a position, a removal or a score change, and the ranking has no affiliate links.