AI Humanizer BenchmarkNovember 1, 2026
AI Humanizer Benchmark: November 2026 AI Humanizer Rankings
The November cycle is in: 10 AI humanizers, each rewriting the same 33 freshly generated texts, every output scored by 7 commercial AI detectors plus meaning preservation and readability. That is 330 rewrites and 2,310 detector verdicts. This month 4 rewrites, 1.2% of the field's output, were judged human by all seven detectors at once, down from 110 in October. The reason is the panel, not the tools.
Two detectors changed. Pangram joined the panel and Winston AI left it, and Originality.ai now scores with its 5.0 model. Both Pangram and the new Originality are far stricter than what they replaced: Pangram judged 6% of all rewrites human, Originality 5%. With two such detectors in a seven-member panel, the median verdict that sets a test's bypass score moved for every tool. Averaged across the field, bypass fell from 70.9% in October to 51.5%. The methodology version is 1.2.0 and the changelog records what moved.
The roster changed too. Rephrasy joins. WriteHuman and AIHumanize could not be exercised this cycle, so they are not in it.
CleverHumanizer won the cycle at 79.33 out of 100 with the field's highest bypass rate, 90%. UndetectedGPT is second at 76.03 and again produced the most readable output at 84%. SmartHumanizer preserved meaning best at 83%. Scores range from 79.33 down to 33.71. The leaderboard has the full table, with every sub-score and per-tool data to download.
November 2026 leaderboard
Full interactive table| # | Humanizer | Overall | Bypass | Meaning | Read. | Pen. |
|---|---|---|---|---|---|---|
| 1▲ 7 | 79.33 | 89.6 | 63.7 | 79.5 | None | |
| 2▼ 1 |  UndetectedGPT | 76.03 | 70.7 | 79.0 | 83.7 | None |
| 3– | 72.97 | 80.0 | 69.6 | 75.3 | −3.0 ×2 | |
| 4▼ 1 | 72.05 | 69.8 | 71.5 | 78.4 | None | |
| 5▲ 1 | 69.41 | 55.2 | 78.9 | 78.3 | None | |
| 6▼ 2 | 67.41 | 48.7 | 82.6 | 81.6 | None | |
| 7▼ 5 | 60.04 | 56.0 | 81.2 | 81.0 | −10.0 | |
| 8▲ 1 | 52.87 | 38.2 | 65.1 | 81.5 | −5.0 ×2 | |
| 9▲ 1 | 44.65 | 3.3 | 69.8 | 82.2 | −2.0 ×2 | |
| 10▲ 1 | 33.71 | 3.5 | 73.3 | 76.5 | −13.0 ×2 |
Overall, bypass, meaning and readability scores run from 0 to 100, and higher is better. The penalty column shows points taken off for output-quality problems. They are already included in the overall score.
From eighth to first
CleverHumanizer finished eighth in October at 68.06, with nine penalty points against it. This month it drew no penalties, and its 90% bypass rate held across the whole panel: it was the best tool against five of the seven detectors, including GPTZero at 83% and Copyleaks at 88%. The panel's median judged 30 of its 33 rewrites human, the most of any tool. Its weak point is the one the formula is built to weigh: 64% meaning preservation, the lowest in the field. Bypass carries 42% of the composite; the other 58% is why its lead over UndetectedGPT is 3.3 points and not more.
Where UndetectedGPT's lead went
UndetectedGPT held first in October at 88.12 on 96% bypass. In November its bypass is 71%. Against the five detectors that stayed in the panel it is close to where it was: Grammarly 93%, QuillBot 92%, Copyleaks 76%, ZeroGPT 71% and GPTZero 61%. Against the two new instruments it is not: 8% on Originality 5.0, against 97% on the old Originality tool, and 6% on Pangram. Its quality scores are unchanged, 79% meaning and the best readability in the field, which is what keeps it second. The same pattern, strong on the five and weak on the two, runs through most of the board.
Rephrasy debuts third
Rephrasy scored 72.97 on its first cycle with the second-best bypass rate, 80%. The panel's median judged 29 of its 33 rewrites human, and two of its rewrites passed all seven detectors, the most of any tool this month. It was also the best tool against Pangram, at 17%. Three penalty points, for one severe meaning drift and two length-inflation triggers, separate it from a higher placing. Its output runs 1.04 times the input on average.
HIX Bypass: from second to seventh
HIX Bypass was second in October at 83.87. It is seventh at 60.04. Part of that is the panel: its bypass went from 93% to 56%. The larger part is length. Its rewrites ran 1.44 times the input on average and triggered the length-inflation penalty 19 times, which reaches the 10-point cap. Its meaning preservation, 81%, is the second-best in the field, so this is a tool that says what the original said, at length.
Last place again
Undetectable AI finished last at 33.71 with a bypass rate of 3%: not one of its 33 rewrites was judged human by the panel's median, and 22 of its 231 individual verdicts cleared the midpoint, the same count as October. It drew 29 length-inflation triggers, output running 1.57 times the input, and three severe-drift penalties, for 13 points off. ReHumanize, ninth at 44.65, was also at 3%. As every month, these are results for each tool's default configuration, which is what this benchmark tests.
The detectors now disagree more
Pangram was the strictest detector this cycle, averaging 5.4% human-likelihood across all 330 rewrites, with Originality.ai at 11.0%. Grammarly was the most permissive at 76.4%, then QuillBot at 69.5% and ZeroGPT at 64.3%. GPTZero sat in the middle at 43.0%, Copyleaks at 56.1%. The gap between the strictest and most permissive detector is now 71 points, against 34 in October. A bypass claim proven against one detector says even less than it did. The detector pages rank every tool against each detector separately.
Business email is still the hardest category
Averaged across all tools, the panel's median judged business emails human 25.6% of the time, against 64.2% for academic essays. Application essays were next hardest at 37.3%. SmartHumanizer, with the best meaning preservation in the field, passed 5% of its business emails. CleverHumanizer passed 83% of its own and Rephrasy 57%; no other tool reached a third. Short, formulaic text gives a rewriter little room to vary structure and vocabulary. If your use case is short-form professional text, the per-category views are the ranking you need.
Did the source model matter?
Averaged across all tools, rewrites of texts written by Gemini 3.5 Flash were judged human 55.7% of the time, the highest of the six models, with Claude Sonnet 5 and GPT-5.6 Luna within a point of it. Rewrites of GPT-5.6 Terra texts were judged human 43.0% of the time, the lowest. In October the lowest was Gemini 2.5 Flash. Each model writes only five or six texts a cycle, so the ordering moves month to month; the spread, about 13 points, is what to watch.
Check any of this yourself
Before this cycle ran, we published a SHA-256 hash of the secret value that picks the prompts, together with the list of models that would write them. That value is now public, and anyone can use it to rebuild the prompt set and the model assigned to each prompt as part of the standard check. The inputs, outputs and detector verdicts for all 330 tests are on the evidence pages and in the public repository. From this cycle the per-tool download files carry their cycle, settings and detector verdicts in every row, and the leaderboard names the readability and embedding models. One command in the repository rebuilds the whole leaderboard from the raw files. If a number in this post does not match what you get, please dispute it.
The December cycle runs on freshly generated texts next month. If you build a humanizer that should be on this board, submit it.