November 2026 AI Humanizer Rankings
As of November 2026, the top-ranked AI humanizer on HumanizerBench is WriteHuman, scoring 70.11 out of 100.
WriteHuman finished first in November at 70.11, ahead of StealthGPT at 69.47 and Clever AI Humanizer at 68.48. We tested sixteen tools this month, two of them for the first time: Ryne AI came in eighth and Rephrasy came last.
The biggest change this month wasn’t one of the humanizers. It was Originality.ai.
| # | Humanizer | Overall | Bypass | Meaning | Read. | Pen. |
|---|---|---|---|---|---|---|
1
-
| 70.11 | 81.1 | 59.2 | 55.4 | −1.0Penalties applied
| |
2
-
| 69.47 | 66.0 | 67.3 | 77.7 | −1.0Penalties applied
| |
3
▲ 1 | 68.48 | 66.7 | 65.8 | 60.9 | None | |
4
▲ 1 | 65.18 | 57.0 | 71.8 | 54.9 | None | |
5
▲ 7 | 64.35 | 91.3 | 52.0 | 42.5 | −7.0Penalties applied
| |
6
▼ 3 | 64.24 | 64.5 | 78.1 | 61.7 | −7.0×2Penalties applied
| |
7
▲ 1 | 62.20 | 52.2 | 75.3 | 39.9 | None | |
8
-
| 61.24 | 45.9 | 77.1 | 47.3 | None | |
9
▲ 2 | 61.05 | 46.2 | 77.5 | 48.6 | None | |
10
▼ 4 | 60.62 | 47.3 | 72.7 | 58.2 | −1.0Penalties applied
| |
11
▼ 4 | 59.57 | 41.5 | 72.9 | 72.6 | −2.0Penalties applied
| |
12
▲ 2 | 58.95 | 31.0 | 86.6 | 67.4 | −2.0Penalties applied
| |
13
-
| 58.38 | 55.2 | 77.5 | 69.3 | −10.0Penalties applied
| |
14
▼ 5 | 57.47 | 53.4 | 77.1 | 68.1 | −10.0Penalties applied
| |
15
▼ 5 | 54.95 | 54.4 | 70.7 | 63.8 | −10.0Penalties applied
| |
16
-
| 52.50 | 73.9 | 44.9 | 52.7 | −11.0×2Penalties applied
|
Originality’s new model catches almost everything
In late October, Originality.ai launched two new detection models, 5.0 and a slower, more expensive 5.0 Max. They replace AI Allowance, the model we had been testing with, which Originality has now retired. We test with 5.0.
The new models work differently. AI Allowance was set to tolerate some AI help before it flagged a text. 5.0 doesn’t give a yes-or-no verdict. It estimates how much of a document involves AI and files it as Likely Original (under 10 percent AI), Mixed Writing, or Likely AI (over 80 percent). Originality says it built the new models for writing that has been edited, rewritten or run through an AI humanizer, which is exactly what we test.
The difference is huge. In October, Originality marked 89 percent of the humanized texts we tested as human. In November it marked 10 percent. On a typical text this month, it flagged about 90 percent of the writing as AI.
That matches Originality’s own testing. Its launch post reports rescanning the 420 humanized texts from our October results with the new model, and says 5.0 flagged 89.5 percent of them.
The humanizers didn’t get worse this month. The other five detectors scored November’s texts about the same as October’s. But the bypass score, which is worth 42 of the 100 points, is the average of all six detectors, so Originality’s change alone took around 5 points off a typical tool’s overall score. Thirteen of the fourteen tools we tested in both months scored lower than in October.
Only three tools held up against the new model. Originality rated 24 of Phrasly’s 30 texts as human, 9 of WriteHuman’s and 7 of StealthGPT’s. No other tool had more than 3.
Pangram updated too, and now it has company
Pangram also released a new model in October. Unlike Originality’s, it didn’t change much overall. Pangram was already the strictest detector we test, and it still rated every text from eight of the sixteen tools as AI.
It did reshuffle the tools that beat it. Pangram rated 20 of WriteHuman’s texts as human, up from 15 last month. Clever AI Humanizer went the other way, from 8 to 1. Phrasly (25), StealthGPT (19) and newcomer Rephrasy (14) were the only other tools that beat it with any regularity.
Only four tools had any text rated human by both Pangram and Originality: Phrasly (20 of 30), WriteHuman (8), StealthGPT (6) and Undetectable.ai (1).
Phrasly climbed, but read what it wrote
Phrasly jumped from twelfth to fifth, the only returning tool to score higher than last month. It got there on detection alone. Its bypass score of 91 percent was the best on the board by a wide margin, and it was the one tool Originality’s new model mostly rated as human.
The writing is bad. Phrasly was second-worst on the board for readability and second-worst for keeping the original meaning, and seven of its texts drifted far enough from what they were given to cost it a penalty point each.
StealthGPT writes the best, but slipped on detection
StealthGPT finished second again. It’s the best writer in the field: it had the top readability score for the second month running, and it kept more of the original meaning than it did in October.
Detection is where it slipped. Its bypass score fell from 88 percent to 66 percent, the biggest drop of any returning tool. Most of that is Originality, but GPTZero also caught it a lot more often than last month. It was uneven across writing types too, scoring 80 percent on blog posts and 46 percent on application essays.
Clever AI Humanizer moves up to third
Clever AI Humanizer, which is free to use, moved up a place to third. It took no penalties this month after losing four points to meaning drift in October, and its scores were steady across every type of writing.
Its weak spot is the strict detectors. Pangram rated 1 of its texts as human this month, down from 8, and Originality’s new model rated only one as human too. Its bypass score fell from 80 percent to 67 percent.
Why WriteHuman finished first
WriteHuman did well on both halves of the score. Its bypass score of 81 percent was second only to Phrasly’s, and it kept its writing intact while getting there: one penalty all month, against Phrasly’s seven.
It was close to perfect against GPTZero, Copyleaks and Winston. Pangram rated 20 of its 30 texts as human, up from 15 in October, and Originality’s new model rated 9 as human, more than any tool except Phrasly. Its rewrites change more of the wording than the most conservative tools do, so its meaning score sits below theirs, but only one of its 30 texts drifted far enough to cost a point.
Ryne AI and Rephrasy are new
Ryne AI came in eighth at 61.24 in its first month. It’s careful with your text: it kept the meaning well, it was steady across writing types and it took no penalties. It just doesn’t beat the strict detectors. GPTZero gave 26 of its 30 texts a flat zero, and neither Pangram nor Originality rated a single one as human.
Rephrasy is the reverse, and it finished last at 52.50. Its bypass score of 74 percent was third-best on the board, and Pangram rated 14 of its texts as human, which most tools can’t manage at all. But it kept less of the original meaning than any other tool. A third of its texts drifted far enough to be penalised, and penalties took 11 points off its score, more than anyone else’s.
HIX Bypass started padding
We take a point off for every text that comes back more than 40 percent longer than what went in, up to 10 points. Padding is a cheap way to dilute the AI signal, and a much longer text isn’t a rewrite of what you wrote.
Last month none of HIX Bypass’s texts crossed that line. This month 13 did, with outputs averaging about 500 words from inputs of about 360. GPTZero did rate it as human far more often, up from 11 percent to 57, which fits what we’ve seen before: longer output tends to fool detectors a little more. It didn’t pay off. HIX Bypass hit the 10-point cap and fell from ninth to fourteenth.
Humbot and Walter Writes also hit the cap, with 13 and 10 padded texts. Walter Writes fell from tenth to fifteenth and lost more points than any other tool this month.
Cover letters are still the hardest
Cover letters were the hardest thing to humanize for the second month in a row, and the only kind of writing where the average bypass score was below half. That’s true with or without Originality. They follow a tight, familiar formula, and that seems to leave humanizers little room to work.
News articles were the surprise. They were the easiest kind of writing in October and the second-hardest this month, and that isn’t just Originality: GPTZero and Copyleaks caught a lot more of them too. We write new source texts every cycle, so it could be down to this month’s particular articles. We’ll see whether it holds in December. Discussion posts were the easiest this month.
What November means
November was a detector story more than a humanizer story. Originality’s new model reset everyone’s scores, and the tools that did best against it, Phrasly, WriteHuman and StealthGPT, are the same three that did best against Pangram last month. Tools that already handle the strictest detector were ready for the next one.
If you’re choosing a tool, the top of the board is a good place to start. WriteHuman, StealthGPT and Clever AI Humanizer all scored well on detection while keeping the writing in good shape, and that balance is what the overall score rewards. If one detector matters most to you, our detector pages rank every tool against each one.
Detectors update often, and the best humanizers keep up with them. We’ll test every tool against the newest version of each detector again in December.