AI Humanizer BenchmarkDetectors / Pangram
Best for Pangram
In the November 2026 cycle, Pangram scored 330 outputs from the 10 humanizers in the benchmark and caught 311 (94.2%), the 2nd-most of the seven detectors. The test set has no human-written text, so false positives are not measured here.
Highest average Pangram human-likelihood this cycle: Rephrasy (17.2), StealthGPT (16.9) and CleverHumanizer (8.8).
An AI text detector used in education and publishing that flags AI-written and AI-assisted passages, with strict verdicts on polished prose.
How we read its score: Reports a 0-1 AI likelihood; inverted to human = 1 - ai.
10 humanizers × 33 prompts = 330 outputs · run November 1, 2026 · methodology v1.2.0
Humanizer ranking against Pangram
Sorted by the average human-likelihood Pangram gave each tool's 33 outputs: its bypass rate against Pangram. No tool had 80% or more of its outputs passed by Pangram. The top three had 6, 6 and 3 of 33 passed.
| Rank | Humanizer | Penalties | Last tested | ||
|---|---|---|---|---|---|
| 1 | 17.2 | 72.97 | -3.0 | ||
| 2 | 16.9 | 52.87 | -5.0 | ||
| 3 | 8.8 | 79.33 | none | ||
| 4 |  UndetectedGPTundetectedgpt.ai | 6.2 | 76.03 | none | |
| 5 | 5.3 | 69.41 | none | ||
| 6 | 0.0 | 72.05 | none | ||
| 7 | 0.0 | 67.41 | none | ||
| 8 | 0.0 | 60.04 | -10.0 | ||
| 9 | 0.0 | 44.65 | -2.0 | ||
| 10 | 0.0 | 33.71 | -13.0 |
10 humanizers · 7 detectors · scores out of 100, higher is better. Click a column to sort; hover a penalty for the breakdown.
Pass rate on Pangram vs the other detectors
Seven humanizers passed Pangram at least 30 percentage points less often than the other six detectors in November 2026: CleverHumanizer, GPTinf, UndetectedGPT, HIX Bypass, Rephrasy, SmartHumanizer and SuperHumanizer. They do well elsewhere but poorly against Pangram specifically.
| Humanizer | Passed Pangram | Passed the other 6 | Difference |
|---|---|---|---|
| 9% 3/33 | 78% 155/198 | −69.2 | |
| 0% 0/33 | 66% 131/198 | −66.2 | |
| UndetectedGPT | 6% 2/33 | 71% 141/198 | −65.2 |
| 0% 0/33 | 63% 124/198 | −62.6 | |
| 18% 6/33 | 78% 155/198 | −60.1 | |
| 0% 0/33 | 57% 112/198 | −56.6 | |
| 6% 2/33 | 62% 123/198 | −56.1 | |
| 18% 6/33 | 41% 81/198 | −22.7 | |
| 0% 0/33 | 19% 38/198 | −19.2 | |
| 0% 0/33 | 11% 22/198 | −11.1 |
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 Pangram minus the pooled rate, in percentage points; negative means Pangram was harder for that tool.
Pangram compared with the other AI detectors
All seven detectors scored the same 330 outputs in November 2026. Pangram gave the same verdict as Originality.ai most often (92.4% of outputs) and as Grammarly least often (24.8%).
| Detector | Caught | Catch rate | Unique catches | Same verdict as Pangram |
|---|---|---|---|---|
| 314 of 330 | 95.2% | 10 | 92.4% | |
| 311 of 330 | 94.2% | 7 | – | |
| 187 of 330 | 56.7% | 0 | 60.6% | |
| 144 of 330 | 43.6% | 0 | 48.8% | |
| 99 of 330 | 30.0% | 1 | 34.5% | |
| 89 of 330 | 27.0% | 0 | 32.1% | |
| 65 of 330 | 19.7% | 0 | 24.8% |
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 Pangram both caught or both passed the output.
Outputs only Pangram caught
In November 2026, Pangram caught 7 outputs that all six other detectors passed: 63.6% of the 11 outputs the others let through. Largest score gap first.
| Prompt | Humanizer | Pangram score | Lowest other score |
|---|---|---|---|
| Product description · Gemini 3.5 Flash | CleverHumanizer | 0.00 | 0.93 |
| Argumentative essay · Claude Haiku 4.5 | CleverHumanizer | 0.00 | 0.92 |
| News article · Gemini 3.5 Flash | CleverHumanizer | 0.00 | 0.92 |
| Lit review · GPT-5.6 Terra | Rephrasy | 0.00 | 0.85 |
| Landing copy · Claude Sonnet 5 | Rephrasy | 0.00 | 0.61 |
| Business email · Gemini 2.5 Flash | Rephrasy | 0.00 | 0.60 |
| Argumentative essay · Gemini 2.5 Flash | Rephrasy | 0.00 | 0.57 |
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.
Pangram catch rate by writing type
In November 2026, Pangram caught business email outputs most often (100%) and marketing copy outputs least often (85%).
| Writing type | Pangram | Other 6 detectors |
|---|---|---|
| Academic essay | 93% 56/60 | 39% 142/360 |
| Application essay | 98% 59/60 | 57% 204/360 |
| Blog post | 95% 57/60 | 48% 171/360 |
| Business email | 100% 30/30 | 55% 99/180 |
| Marketing copy | 85% 51/60 | 39% 142/360 |
| Discussion board | 97% 29/30 | 38% 69/180 |
| News article | 97% 29/30 | 39% 71/180 |
The last column pools the other six detectors' verdicts on the same outputs. By prompt, Pangram caught the most tools on Argumentative essay · GPT-5.6 Terra (10 of 10) and the fewest on Product description · Gemini 3.5 Flash (6 of 10).
Using these results to pick a humanizer
The ranking on this page uses Pangram 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 Pangram 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 Pangram, and compare it with your original for changed facts or dropped citations.
Frequently asked questions about Pangram
Can Pangram detect humanized AI text?
Some of it. In November 2026, Pangram caught 311 of the 330 humanized outputs it scored (94.2%), the 2nd-most of the seven detectors in the benchmark. The test set has no human-written text, so how often Pangram flags human writing is not measured here.
Can Pangram be bypassed?
Rarely. In November 2026, 19 of the 330 humanized outputs (5.8%) passed Pangram as human. No humanizer had at least 80% of their outputs passed.
What is the best AI humanizer for Pangram?
By Pangram alone, Rephrasy: it had the highest average human-likelihood from Pangram in November 2026 (17.2 out of 100), and Pangram caught 27 of its 33 outputs. The leaderboard ranks tools on all seven detectors plus meaning and readability.
Which writing types does Pangram catch most often?
Business email outputs, at 100% in November 2026. Marketing copy outputs were caught least often, at 85%.
How do you test Pangram?
Each month, every humanizer rewrites the same 33 AI-written texts on its default settings, and Pangram 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 November 2026 cycle. Each row has the text Pangram scored and the score it returned. To check a result, paste a row's text into Pangram 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).