AI Humanizer BenchmarkDetectors / Copyleaks
Best for Copyleaks
In the October 2026 cycle, Copyleaks scored 363 outputs from the 11 humanizers in the benchmark and caught 145 (39.9%), the 2nd-most of the seven detectors. The test set has no human-written text, so false positives are not measured here.
Highest average Copyleaks human-likelihood this cycle: WriteHuman (93.9), HIX Bypass (81.8) and UndetectedGPT (78.8).
An AI and plagiarism detection platform used by universities and businesses, where a flag can carry formal consequences.
How we read its score: Reports AI probability; inverted (human = 1 - ai).
11 humanizers × 33 prompts = 363 outputs · run October 1, 2026 · methodology v1.1.0
Humanizer ranking against Copyleaks
Sorted by the average human-likelihood Copyleaks gave each tool's 33 outputs: its bypass rate against Copyleaks. Two of the 11 tools had 80% or more of their outputs passed by Copyleaks. The top three had 31, 27 and 26 of 33 passed. In September 2026, AI Humanize ranked first.
| Rank | Humanizer | Penalties | Last tested | ||
|---|---|---|---|---|---|
| 1 | 93.9 | 75.06 | -4.0 | ||
| 2 | 81.8 | 83.87 | none | ||
| 3 |  UndetectedGPTundetectedgpt.ai | 78.8 | 88.12 | none | |
| 4 | 72.7 | 80.20 | none | ||
| 5 | 72.7 | 78.33 | -1.0 | ||
| 6 | 69.7 | 68.06 | -9.0 | ||
| 7 | 63.6 | 78.23 | -1.0 | ||
| 8 | 60.6 | 64.52 | -4.0 | ||
| 9 | 57.6 | 80.74 | none | ||
| 10 | 4.7 | 41.60 | -4.0 | ||
| 11 | 1.5 | 33.60 | -12.0 |
11 humanizers · 7 detectors · scores out of 100, higher is better. Click a column to sort; hover a penalty for the breakdown.
Pass rate on Copyleaks vs the other detectors
In October 2026, every humanizer's pass rate on Copyleaks was within 30 points of its rate on the other detectors.
| Humanizer | Passed Copyleaks | Passed the other 6 | Difference | Cycles 30+ pts lower |
|---|---|---|---|---|
| 58% 19/33 | 86% 171/198 | −28.8 | 0 of 2 | |
| 70% 23/33 | 85% 168/198 | −15.2 | 1 of 2 | |
| 64% 21/33 | 78% 155/198 | −14.6 | 0 of 2 | |
| UndetectedGPT | 79% 26/33 | 92% 183/198 | −13.6 | 1 of 2 |
| 6% 2/33 | 19% 38/198 | −13.1 | 0 of 2 | |
| 73% 24/33 | 84% 167/198 | −11.6 | 0 of 2 | |
| 3% 1/33 | 11% 21/198 | −7.6 | 0 of 2 | |
| 94% 31/33 | 97% 193/198 | −3.5 | 0 of 2 | |
| 61% 20/33 | 62% 122/198 | −1.0 | 0 of 2 | |
| 73% 24/33 | 72% 143/198 | +0.5 | 1 of 2 | |
| 82% 27/33 | 79% 157/198 | +2.5 | 0 of 2 |
Passed means a human-likelihood of 0.50 or higher. The other six detectors' verdicts on the same outputs are pooled. Difference is the pass rate on Copyleaks minus the pooled rate, in percentage points; negative means Copyleaks was harder for that tool.
Copyleaks compared with the other AI detectors
All seven detectors scored the same 363 outputs in October 2026. Copyleaks gave the same verdict as Winston AI most often (79.1% of outputs) and as GPTZero least often (68.0%).
| Detector | Caught | Catch rate | Change since Sep 2026 | Unique catches | Same verdict as Copyleaks |
|---|---|---|---|---|---|
| 203 of 363 | 55.9% | +11.3 | 64 | 68.0% | |
| 145 of 363 | 39.9% | −19.6 | 13 | – | |
| 125 of 363 | 34.4% | −12.4 | 8 | 79.1% | |
| 99 of 363 | 27.3% | −30.3 | 0 | 78.5% | |
| 95 of 363 | 26.2% | +1.9 | 7 | 79.1% | |
| 71 of 363 | 19.6% | −1.7 | 1 | 75.2% | |
| 67 of 363 | 18.5% | +1.4 | 1 | 76.3% |
Change is in percentage points of catch rate. Unique catches are outputs a detector caught while every other detector passed them. Same verdict is the share of outputs where the detector and Copyleaks both caught or both passed the output.
Outputs only Copyleaks caught
In October 2026, Copyleaks caught 13 outputs that all six other detectors passed: 10.6% of the 123 outputs the others let through. Largest score gap first; the top 10 are shown.
| Prompt | Humanizer | Copyleaks score | Lowest other score |
|---|---|---|---|
| Lit review · Claude Sonnet 5 | CleverHumanizer | 0.00 | 1.00 |
| Lit review · Claude Sonnet 5 | StealthGPT | 0.00 | 0.99 |
| News article · Claude Haiku 4.5 | UndetectedGPT | 0.00 | 0.95 |
| Landing copy · Gemini 2.5 Flash | CleverHumanizer | 0.00 | 0.88 |
| Product description · Gemini 2.5 Flash | GPTinf | 0.00 | 0.83 |
| Lit review · Gemini 2.5 Flash | CleverHumanizer | 0.00 | 0.83 |
| Business email · Gemini 2.5 Flash | UndetectedGPT | 0.00 | 0.80 |
| Cover letter · Claude Sonnet 5 | CleverHumanizer | 0.00 | 0.79 |
| How-to blog · Gemini 3.5 Flash | GPTinf | 0.00 | 0.77 |
| How-to blog · Gemini 3.5 Flash | UndetectedGPT | 0.00 | 0.76 |
Scores are human-likelihood from 0 to 1. Every other detector scored these outputs 0.50 or higher; the last column is the lowest of those scores. Each prompt opens its evidence page at that humanizer's output.
Copyleaks catch rate by writing type
In October 2026, Copyleaks caught business email outputs most often (91%) and discussion board outputs least often (24%). Academic essay was its least-caught type in both cycles.
| Writing type | Copyleaks | Other 6 detectors |
|---|---|---|
| Academic essay | 24% 16/66 | 22% 87/396 |
| Application essay | 52% 34/66 | 39% 155/396 |
| Blog post | 36% 24/66 | 32% 126/396 |
| Business email | 91% 30/33 | 41% 82/198 |
| Marketing copy | 33% 22/66 | 29% 115/396 |
| Discussion board | 24% 8/33 | 27% 53/198 |
| News article | 33% 11/33 | 21% 42/198 |
The last column pools the other six detectors' verdicts on the same outputs. By prompt, Copyleaks caught the most tools on Cover letter · GPT-5.6 Luna (10 of 11) and the fewest on Product description · Claude Sonnet 5 (1 of 11).
Copyleaks by month
| Month | Humanizers | Caught | Catch rate | Avg human-likelihood | Rank by catches |
|---|---|---|---|---|---|
| October 2026 | 11 | 145 of 363 | 39.9% | 59.8 | 2 of 7 |
| September 2026 | 11 | 216 of 363 | 59.5% | 40.4 | 1 of 7 |
Rank 1 is the detector that caught the most outputs that month.
Using these results to pick a humanizer
The ranking on this page uses Copyleaks only. It does not account for the other six detectors, for how well a rewrite keeps the original meaning, or for readability. A tool can pass Copyleaks while changing facts or reading poorly. The Overall column above and the leaderboard cover those.
Results can shift between cycles as detectors and humanizers update, so check which cycle a figure comes from, and each tool's history on its own page.
Before relying on a tool, run a sample of your own text through it, check the output in Copyleaks, and compare it with your original for changed facts or dropped citations.
Frequently asked questions about Copyleaks
Can Copyleaks detect humanized AI text?
Some of it. In October 2026, Copyleaks caught 145 of the 363 humanized outputs it scored (39.9%), the 2nd-most of the seven detectors in the benchmark. The test set has no human-written text, so how often Copyleaks flags human writing is not measured here.
Can Copyleaks be bypassed?
Often. In October 2026, 218 of the 363 humanized outputs (60.1%) passed Copyleaks as human. Two of the 11 humanizers had at least 80% of their outputs passed.
What is the best AI humanizer for Copyleaks?
By Copyleaks alone, WriteHuman: it had the highest average human-likelihood from Copyleaks in October 2026 (93.9 out of 100), and Copyleaks caught 2 of its 33 outputs. The leaderboard ranks tools on all seven detectors plus meaning and readability.
Which writing types does Copyleaks catch most often?
Business email outputs, at 91% in October 2026. Discussion board outputs were caught least often, at 24%.
How do you test Copyleaks?
Each month, every humanizer rewrites the same 33 AI-written texts on its default settings, and Copyleaks scores every rewrite. Each result is converted to a human-likelihood from 0 to 1, as described above. A score below 0.50 counts as caught.
Raw test data
A JSONL file with one row per test from the October 2026 cycle. Each row has the text Copyleaks scored and the score it returned. To check a result, paste a row's text into Copyleaks and compare.
Other detectors
Related
How scores are computed
See the full methodologyOverall score formula
How often the rewritten text is classified as human by the detector panel.
How closely the rewrite preserves the original text's meaning.
Whether the output is clear, fluent, natural writing.
Whether performance holds across all seven writing types or varies widely between them.
Penalties (right) subtract from the composite, and the result never goes below zero.
Possible penalties
- Identical to input
The text came back essentially unchanged; no meaningful rewriting occurred.
-2.0-10.0 - Refusal
The tool declined to process the text, typically due to its own content filter.
-1.0-10.0 - Meaning drift
The rewrite departed substantially from the original's meaning.
-1.0-10.0 - Length inflation
The output ran far longer than the input, diluting the AI signal with added words rather than removing it.
-1.0-10.0 - Length deflation
The output ran far shorter than the input; content was dropped rather than rephrased.
-1.0-10.0
Maximum possible total penalty: -50.0 (off 100).