AI Detectors in the Benchmark

Each monthly cycle, every humanizer in the benchmark rewrites the same 33 AI-generated texts, and seven AI detectors score each rewrite. Across the 363 humanized outputs of October 2026, catch rates ranged from 18.5% (Grammarly) to 55.9% (GPTZero). The test set contains only humanized AI text: it measures how often detectors miss AI writing, not how often they flag human writing by mistake.

GPTZero

All October 2026 outputs Leaderboard Raw scores on GitHub

11 humanizers × 33 prompts = 363 outputs · scored by 7 detectors · run October 1, 2026 · methodology v1.1.0

How often each detector catches humanized text

October 2026 cycle, sorted by catch rate. Every detector scored the same 363 outputs from 11 humanizers.

Detector	Caught	Catch rate	Avg human-likelihood	Change since Sep 2026	Unique catches
GPTZerogptzero.me	203 of 363	55.9%	43.6	+11.3	64
Copyleakscopyleaks.com	145 of 363	39.9%	59.8	−19.6	13
Winston AIgowinston.ai	125 of 363	34.4%	64.9	−12.4	8
Originality.aioriginality.ai	99 of 363	27.3%	72.2	−30.3	0
QuillBotquillbot.com	95 of 363	26.2%	72.0	+1.9	7
ZeroGPTzerogpt.com	71 of 363	19.6%	76.2	−1.7	1
Grammarlygrammarly.com	67 of 363	18.5%	77.7	+1.4	1

A detector catches an output when it gives a human-likelihood below 0.50 (scale 0 to 1). Avg human-likelihood is the mean of its scores, shown out of 100. Change is the catch-rate difference from September 2026, in percentage points. Unique catches are outputs this detector caught and every other detector passed.

Largest change from September 2026: Originality.ai's catch rate fell from 57.6% to 27.3%. It dropped for 10 of the 11 humanizers tested in both cycles.

Results by writing type

Catch rate per detector for each of the seven writing types, October 2026. Darker cells mean a higher catch rate.

Detector	Academic essay	Application essay	Blog post	Business email	Marketing copy	Discussion board	News article
GPTZerogptzero.me	47% 31/66	59% 39/66	50% 33/66	82% 27/33	52% 34/66	61% 20/33	58% 19/33
Copyleakscopyleaks.com	24% 16/66	52% 34/66	36% 24/66	91% 30/33	33% 22/66	24% 8/33	33% 11/33
Winston AIgowinston.ai	23% 15/66	33% 22/66	45% 30/66	58% 19/33	33% 22/66	24% 8/33	27% 9/33
Originality.aioriginality.ai	18% 12/66	27% 18/66	36% 24/66	24% 8/33	35% 23/66	21% 7/33	21% 7/33
QuillBotquillbot.com	21% 14/66	50% 33/66	18% 12/66	33% 11/33	21% 14/66	18% 6/33	15% 5/33
ZeroGPTzerogpt.com	12% 8/66	39% 26/66	20% 13/66	15% 5/33	20% 13/66	18% 6/33	0% 0/33
Grammarlygrammarly.com	11% 7/66	26% 17/66	21% 14/66	36% 12/33	14% 9/66	18% 6/33	6% 2/33

Sample sizes differ: business email, discussion board and news article have 3 prompts per cycle (33 outputs per detector); the other types have 6 (66 outputs).

Month-by-month results

Catch rate per detector in each published cycle, newest first. Grammarly had the lowest in both cycles.

Detector	October 2026	September 2026
GPTZerogptzero.me	55.9% 203/363	44.6% 162/363
Copyleakscopyleaks.com	39.9% 145/363	59.5% 216/363
Winston AIgowinston.ai	34.4% 125/363	46.8% 170/363
Originality.aioriginality.ai	27.3% 99/363	57.6% 209/363
QuillBotquillbot.com	26.2% 95/363	24.2% 88/363
ZeroGPTzerogpt.com	19.6% 71/363	21.2% 77/363
Grammarlygrammarly.com	18.5% 67/363	17.1% 62/363

Agreement between detectors

In October 2026, every detector passed 110 of the 363 outputs (30.3%) and every detector caught 34. The remaining 219 got mixed verdicts. Pairwise agreement was highest between ZeroGPT and QuillBot (87.9% of outputs with the same verdict) and lowest between GPTZero and Grammarly (57.6%). Browse the per-prompt verdicts.

The detector panel

GPTZero

gptzero.me

One of the earliest AI text detectors, widely used in education to check whether student writing was generated with AI.

Highest average human-likelihood from GPTZero: WriteHuman at 100.0.

GPTZero bypass ranking

Copyleaks

copyleaks.com

An AI and plagiarism detection platform used by universities and businesses, where a flag can carry formal consequences.

Highest average human-likelihood from Copyleaks: WriteHuman at 93.9.

Copyleaks bypass ranking

Winston AI

gowinston.ai

An AI detector and plagiarism checker that scores documents for AI content, plagiarism and readability.

Highest average human-likelihood from Winston AI: WriteHuman at 94.8.

Winston AI bypass ranking

Originality.ai

originality.ai

A paid AI and plagiarism checker used by publishers, SEO teams and content teams before content goes live.

Highest average human-likelihood from Originality.ai: WriteHuman at 99.8.

Originality.ai bypass ranking

QuillBot

quillbot.com

The AI detector inside the QuillBot writing suite, widely used by students.

Highest average human-likelihood from QuillBot: HIX Bypass at 93.4.

QuillBot bypass ranking

ZeroGPT

zerogpt.com

A browser-based AI detector with wide consumer reach, often used for quick checks of suspected AI text.

Highest average human-likelihood from ZeroGPT: HIX Bypass at 94.7.

ZeroGPT bypass ranking

Grammarly

grammarly.com

The AI detector built into Grammarly's writing assistant, where much drafting happens.

Highest average human-likelihood from Grammarly: UndetectedGPT at 96.8.

Grammarly bypass ranking

Frequently asked questions about AI detectors

Which AI detector catches the most humanized text?

GPTZero, in the October 2026 cycle. It caught 203 of 363 humanized outputs (55.9%). Grammarly had the lowest catch rate: 67 of 363 (18.5%).

How well do AI detectors catch humanized text?

Unevenly. In October 2026, catch rates ran from 18.5% to 55.9%, and six of the seven detectors passed more than half of the humanized outputs as human. The test set has no human-written text, so false positives are not measured.

How often do AI detectors give the same verdict?

In October 2026, all seven detectors reached the same verdict on 144 of the 363 outputs (39.7%): 110 passed by every detector and 34 caught by every detector. The closest pair, ZeroGPT and QuillBot, matched on 87.9% of outputs.

What counts as a catch?

A human-likelihood below 0.50 on a 0 to 1 scale. Detectors report results in different formats (an AI probability, a human score or an AI percentage), so each result is converted to this scale first. Each detector's page lists its conversion.

A humanizer's bypass rate against a detector is the average human-likelihood that detector gave its outputs. For the full method, see how we measure bypass rate.