AI detector benchmark. Numbers, not promises.
We run a fixed set of AI-written samples through HumanizerGPT and widely used AI detectors, and publish what happens, including where we fall short.
Sample data — the first public benchmark run is being prepared. The figures below show the format of the report, not measured results.
Pass rate by detector
Share of samples each detector classified as human-written, before and after humanizing in Balanced mode.
Raw vs humanized
% of samples classified as human- Turnitin2% → 91%
- GPTZero4% → 94%
- Originality.ai1% → 86%
- Copyleaks3% → 92%
- ZeroGPT6% → 97%
- Winston AI2% → 89%
October 2026 results
n = 500| Detector | Raw | Humanized | Δ pts | vs last month |
|---|---|---|---|---|
| Turnitin | 2% | 91% | +89 | +2 |
| GPTZero | 4% | 94% | +90 | +1 |
| Originality.ai | 1% | 86% | +85 | −3 |
| Copyleaks | 3% | 92% | +89 | +4 |
| ZeroGPT | 6% | 97% | +91 | 0 |
| Winston AI | 2% | 89% | +87 | +5 |
| Average | 3% | 91.5% | +88.5 |
Pass = the detector classifies the text as human-written. Inputs: GPT-5, Claude, Gemini and DeepSeek drafts across 5 content types.
What the pass rate means
A sample passes a detector when the detector classifies it as human-written using its default settings. The pass rate is the share of samples that pass. We report two numbers for every detector: the pass rate of the raw AI draft, and the pass rate of the same draft after HumanizerGPT.
The gap between the two shows how much the rewrite changes a detector's verdict. The raw number matters too: it shows how easily each detector recognises unedited AI text in the first place.
Detectors we test
- Turnitin — widely used by universities and schools. It has no public API, so its results come from manual checks on a smaller subset.
- GPTZero — popular with teachers and students.
- Originality.ai — common with publishers and SEO agencies.
- Copyleaks — used in education and enterprise.
- ZeroGPT — a free, widely used checker.
- Winston AI — aimed at education and publishing.
What else we measure
Passing a detector is worthless if the text no longer says what you meant. That is why every run also reports meaning similarity between the draft and the rewrite, readability, and processing time. A rewrite that drifts from the original counts against us, even if it passes.
Limits of any benchmark
Detectors change their models without notice, so a result from one month may not hold the next. Results also vary by text type and length: short texts are harder for detectors to judge either way. Treat these numbers as a guide to how the tool behaves at scale, not as a promise about any single document.
Benchmark questions
Are these real benchmark results?
Not yet. The numbers on this page are sample data that show how the report will look. The first public benchmark run is being prepared, and this page will switch to real results when it is published.
Does a high pass rate mean my text will pass Turnitin?
No. No tool can guarantee a detector result. A pass rate describes how a large test set performed on one date; your own text, the detector version and its settings can all change the outcome.
Why publish numbers at all?
Many humanizers claim to bypass every detector without showing evidence. We would rather publish our method and results, including the detectors where we do worst, so you can judge for yourself.
How often is the benchmark updated?
We plan to run the full test set once a month and after every model update, and to keep previous months available for comparison.
Can I get the raw data?
Yes. Once real runs are published, researchers and journalists can request the per-sample results. The methodology page explains how.
Test it on your own text.
Humanize up to 500 words a day for free.