AI Humanizer BenchmarkOctober 1, 2026
AI Humanizer Benchmark: October 2026 AI Humanizer Rankings
The October cycle is in: 11 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 363 rewrites and 2,541 detector verdicts. This month 110 rewrites, 30.3% of the field's output, were judged human by all seven detectors at once. Seven in ten were still caught by at least one.
One thing is new. The source texts are no longer written by a single model: six models from Anthropic, OpenAI and Google wrote this cycle's 33 texts, and the same published commitment that fixes the prompts now also fixes which model writes each one. Every text's model is recorded in the data and shown on its evidence page.
UndetectedGPT won the cycle at 88.12 out of 100 and also produced the field's most readable output at 84%. WriteHuman posted the highest raw bypass rate at 98%, and SmartHumanizer preserved meaning best at 83%. Scores range from 88.12 down to 33.60. The leaderboard has the full table, with every sub-score and per-tool data to download.
October 2026 leaderboard
Full interactive table| # | Humanizer | Overall | Bypass | Meaning | Read. | Pen. |
|---|---|---|---|---|---|---|
| 1– |  UndetectedGPT | 88.12 | 96.0 | 77.4 | 83.6 | None |
| 2▲ 4 | 83.87 | 93.5 | 77.9 | 64.3 | None | |
| 3▲ 1 | 80.74 | 88.7 | 68.3 | 77.7 | None | |
| 4▼ 2 | 80.20 | 78.2 | 82.9 | 77.5 | None | |
| 5– | 78.33 | 89.5 | 66.4 | 72.1 | −1.0 | |
| 6▲ 2 | 78.23 | 83.9 | 75.2 | 71.0 | −1.0 | |
| 7▼ 4 | 75.06 | 97.6 | 53.2 | 71.4 | −4.0 | |
| 8▲ 1 | 68.06 | 85.9 | 65.5 | 76.8 | −9.0 ×2 | |
| 9▼ 2 | 64.52 | 62.8 | 67.7 | 80.6 | −4.0 ×3 | |
| 10– | 41.60 | 1.7 | 68.7 | 82.7 | −4.0 ×2 | |
| 11– | 33.60 | 1.8 | 72.9 | 73.2 | −12.0 ×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.
The tool that beat every detector finished seventh
WriteHuman cleared all seven detectors on 30 of its 33 rewrites, nearly twice as many as any other tool; UndetectedGPT was next with 16. Its bypass rate, 98%, is the best in the field. It still finished seventh, down from third in September, because its rewrites drifted furthest from what the originals said: 53% meaning preservation, the lowest of all eleven tools, with four severe-drift penalties on top.
This is the case the scoring formula exists for, and the second cycle running that it has decided a placing. A rewrite that beats detectors by saying something different has not humanized the text; it has replaced it. Bypass carries 42% of the composite so that the other 58% can stop that strategy from winning.
How HIX Bypass climbed four places with one clean sweep
HIX Bypass moved from sixth to second at 83.87, on 93% bypass and 78% meaning. It did so while clearing all seven detectors on just one rewrite. GPTZero judged a single one of its 33 outputs human, a 3% pass rate, while each of the other six detectors passed at least 82% of them. Because a test's bypass score is the median verdict across the panel, one detector that catches nearly everything costs very little when the rest agree. The same rule would work against a tool that fooled only one detector. Its weak point is readability: 64%, the lowest in the field.
What separates first from the pack
UndetectedGPT's margin, 4.25 points, came from being strong everywhere at once: 96% bypass, 77% meaning, 84% readability, 97% consistency and no penalties. Its lowest category score for bypass was 91%, on business email. Behind HIX Bypass the board is tight. GPTinf took third at 80.74, SmartHumanizer fourth at 80.20, and about 2.5 points cover third through sixth. SmartHumanizer slipped from second with the best meaning score in the field but a 78% bypass rate, the lowest of the top seven. You can put any two of these side by side.
Last place again
Undetectable AI finished last at 33.60 with a bypass rate of 2%: not one of its 33 rewrites was judged human by the panel's median verdict, and 22 of its 231 individual detector verdicts cleared the midpoint. Its meaning preservation, 73%, was fifth-best. It drew 25 length-inflation triggers, capped at the maximum penalty, with output running 1.53 times the input on average. ReHumanize, tenth at 41.60, was also at 2% bypass. As in September, these are results for each tool's default configuration, which is what this benchmark tests.
The detectors still disagree
GPTZero was the strictest detector this cycle, averaging 43.6% human-likelihood across all 363 rewrites; Grammarly was the most permissive at 77.7%. The disagreements are not uniform. UndetectedGPT cleared ZeroGPT and Grammarly on every test but GPTZero on 76% of them. GPTinf cleared Winston AI on 91% but Copyleaks on 58%. HIX Bypass, as above, went from 3% against GPTZero to 100% against Grammarly.
The practical conclusion is the same as last month: a bypass claim proven against one detector says very little. The detector pages rank every tool against each detector separately.
Business email is still the hardest category
Averaged across all tools, business email was humanized successfully 56.6% of the time, against 79.9% for academic essays. Application essays were close behind it at 59.1%. Two tools scored 0% on business email, and StealthGPT, ninth overall, managed 18% on application essays. Short, formulaic text gives a rewriter little room to vary structure and vocabulary. If your use case is short-form professional text, the overall ranking is not the ranking you need; the per-category views are.
Did the source model matter?
A first look says somewhat. Averaged across all tools, rewrites of texts written by Gemini 3.5 Flash were judged human 79.1% of the time, the highest of the six models, and rewrites of Gemini 2.5 Flash texts 66.4%, the lowest. The other four models fell between 67% and 73%. Each model wrote only five or six texts, each from a different prompt, so this is a spread to watch over coming cycles, not a conclusion.
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 363 tests are on the evidence pages and in the public repository. One command there rebuilds the whole leaderboard from the raw files. If a number in this post does not match what you get, please dispute it.
The November cycle runs on freshly generated texts next month. If you build a humanizer that should be on this board, submit it.