AI Detectors We Test Against
Every month each AI humanizer we test rewrites the same set of AI-written passages, and six AI detectors score every rewrite. In October 2026 they caught between 7.1% and 82.9% of the 420 humanized outputs: Pangram caught the most and Winston AI the fewest. We only test humanized AI text, so these figures say nothing about how accurate detectors are on human writing.
14 humanizers · 30 prompts each · 6 detectors · Last tested · Methodology v1.3.0
Catch rate by detector
| Detector | Caught | Rate Catch rate | Avg score Avg human score | vs Sep 2026 Change since Sep 2026 | Only it Only it caught |
|---|---|---|---|---|---|
 Pangram pangram.com | 348 of 420 | 82.9% | 16.9 | n/a | 147 |
 GPTZero gptzero.me | 151 of 420 | 36.0% | 61.3 | −11.2 | 1 |
 ZeroGPT zerogpt.com | 62 of 420 | 14.8% | 76.5 | −3.9 | 0 |
 Copyleaks copyleaks.com | 58 of 420 | 13.8% | 85.9 | −19.1 | 0 |
 Originality.ai originality.ai | 48 of 420 | 11.4% | 87.2 | −12.1 | 1 |
 Winston AI gowinston.ai | 30 of 420 | 7.1% | 91.8 | −28.3 | 1 |
Swipe the table sideways to see every column.
Caught means a human-likelihood score below 0.50 on the detector's 0 to 1 scale. Avg human score is the mean score on a 0 to 100 scale. Change is in points. Only it caught counts outputs no other detector caught, of the 420 all six scored.
Copyleaks caught 13.8% in October 2026, against 32.9% in September 2026. Ten of the 12 humanizers tested in both months passed Copyleaks more often in October 2026 than in September 2026. Winston AI caught 7.1% in October 2026, against 35.4% in September 2026. All 12 humanizers tested in both months passed Winston AI more often in October 2026 than in September 2026.
Catch rate by month
| Detector | Oct 2026 October 2026 | Sep 2026 September 2026 | Aug 2026 August 2026 | Jul 2026 July 2026 | Jun 2026 June 2026 |
|---|---|---|---|---|---|
| 82.9% 348/420 | n/a | n/a | n/a | n/a | |
| 36.0% 151/420 | 47.2% 217/460 | 47.2% 187/396 | 50.6% 217/429 | 45.6% 178/390 | |
| 14.8% 62/420 | 18.7% 86/460 | 19.4% 77/396 | 15.4% 66/429 | 24.1% 94/390 | |
| 13.8% 58/420 | 32.9% 146/444 | 42.7% 169/396 | 41.8% 179/428 | 40.0% 156/390 | |
| 11.4% 48/420 | 23.5% 108/460 | 55.8% 221/396 | 58.5% 251/429 | 53.8% 210/390 | |
| 7.1% 30/420 | 35.4% 157/444 | 36.1% 143/396 | 36.2% 155/428 | 31.3% 122/390 |
Swipe the table sideways to see every column.
The field of humanizers changes a little from month to month, so read each column as that month's test.
Catch rate by writing type
How often each detector caught humanized outputs of each writing type in October 2026.
| Detector | Academic Essay | Application Essay | Blog Post | Marketing Copy | Discussion Board | News Article |
|---|---|---|---|---|---|---|
| 85% 71/84 | 87% 73/84 | 85% 71/84 | 81% 68/84 | 79% 33/42 | 76% 32/42 | |
| 30% 25/84 | 50% 42/84 | 20% 17/84 | 42% 35/84 | 50% 21/42 | 26% 11/42 | |
| 13% 11/84 | 7% 6/84 | 33% 28/84 | 17% 14/84 | 5% 2/42 | 2% 1/42 | |
| 21% 18/84 | 19% 16/84 | 17% 14/84 | 5% 4/84 | 7% 3/42 | 7% 3/42 | |
| 10% 8/84 | 4% 3/84 | 13% 11/84 | 21% 18/84 | 14% 6/42 | 5% 2/42 | |
| 4% 3/84 | 10% 8/84 | 7% 6/84 | 11% 9/84 | 10% 4/42 | 0% 0/42 |
Swipe the table sideways to see every column.
Discussion Board and News Article had 3 prompts this month against 6 for the other types.
How often the detectors agree
Of the 420 outputs all six detectors scored in October 2026, all six passed 68 (16.2%), all six caught 2, and they split on the other 350. Winston AI and Originality.ai agreed most often, giving the same verdict on 86.2% of outputs. Read the verdicts prompt by prompt
The detectors we test
-
GPTZero is one of the earliest AI text detectors, widely used by educators to check whether student writing was generated with AI.
-
Originality.ai is a subscription AI checker used by publishers, SEO teams, and content teams to detect AI writing and plagiarism.
-
Copyleaks is an enterprise AI checker and plagiarism-detection platform used by universities, businesses, and content-moderation teams.
-
Winston AI is an AI detector and plagiarism checker used to review writing for AI-generated content, readability, and originality.
-
ZeroGPT is a free, browser-based AI detector popular for quick checks of suspected AI-generated content.
-
Pangram is an AI detector from Pangram Labs, used by schools, publishers, and online platforms, that labels each passage of a text as human-written, AI-assisted, or AI-generated.
Questions about AI detectors
Which AI detector is hardest to pass?
In October 2026 it was Pangram, which caught 348 of 420 humanized outputs (82.9%). In September 2026 it was GPTZero. In June, July and August 2026 it was Originality.ai.
Are AI detectors accurate on humanized text?
It depends on the detector and the month. Over the five months we have tested them, catch rates on humanized outputs ran from 7.1% (Winston AI in October 2026) to 82.9% (Pangram in October 2026). These are catch rates on humanized AI text only, not on human writing.
Do AI detectors agree with each other?
Often, but not always. In October 2026 all six gave the same verdict on 16.7% of the 420 outputs they all scored. Pangram and Winston AI agreed least often, with the same verdict on 23.3% of outputs.
Which AI detectors does HumanizerBench test?
Six: GPTZero, Originality.ai, Copyleaks, Winston AI, ZeroGPT and Pangram. Every humanized output goes to each of them every month, and every verdict is published in detector-scores.json.