AI humanizer benchmark: the best tools of July 2026
The July 2026 cycle of HumanizerBench tested 13 AI humanizers to measure which is best at bypassing AI detection without degrading the text it is handed. WriteHuman finished first with a composite score of 73.07, ahead of Undetectable AI at 72.17 and Humanize AI Pro at 70.49. The complete table, with every sub-score and penalty for all 13 tools, is on the leaderboard. The order at the top is close, but the ranking is less informative than the sub-scores and penalties that produced it. What the July results show most clearly is the difference between a high score earned by rewriting text well and a high score earned by exploiting how AI detectors work.
Every tool on the board is measured on two properties that frequently pull in opposite directions. The first is detection bypass: the share of commercial detector tests that read a rewrite as human rather than machine-generated. The second is fidelity: whether the rewrite preserved the length and the meaning of the text it was given. A tool can post a strong bypass number while degrading the text in the process, and this cycle contains a precise example of that trade being made.
The composite score weights these properties deliberately. Most of the weight rests on detection bypass, since evading detectors is the function these tools are sold to perform. The remainder is divided between meaning preservation and readability, and a set of penalties then subtracts points for specific distortions: padding, over-trimming, and meaning drift among them. The effect of that structure is that a tool can lead on the property carrying the most weight and still be overtaken once the penalties are applied. This is the pattern at the top of the table this cycle, and it is why the sub-scores and penalties warrant closer attention than the composite alone.
| # | Humanizer | Overall | Bypass | Meaning | Read. | Pen. |
|---|---|---|---|---|---|---|
1
▲ 1 | 73.07 | 81.6 | 72.9 | 56.2 | −1.0Penalties applied
| |
2
▲ 3 | 72.17 | 95.7 | 73.0 | 55.9 | −10.0Penalties applied
| |
3
▲ 6 | 70.49 | 70.4 | 74.3 | 60.3 | None | |
4
▼ 3 | 68.07 | 81.2 | 68.5 | 44.5 | −3.0×2Penalties applied
| |
5
▼ 1 | 66.42 | 70.7 | 75.3 | 43.9 | −1.0Penalties applied
| |
6
▼ 3 | 64.84 | 65.9 | 73.8 | 42.9 | None | |
7
▲ 3 | 62.84 | 80.6 | 64.7 | 61.5 | −8.0×2Penalties applied
| |
8
▼ 1 | 61.52 | 81.6 | 63.0 | 60.6 | −10.0×3Penalties applied
| |
9
▼ 1 | 61.47 | 73.4 | 60.9 | 72.1 | −8.0×3Penalties applied
| |
10
▼ 4 | 59.66 | 76.0 | 69.8 | 55.7 | −10.0×2Penalties applied
| |
11
-
| 54.13 | 42.3 | 73.2 | 67.1 | −5.0×3Penalties applied
| |
12
-
| 53.38 | 0.0 | 94.5 | 82.1 | None | |
13
-
| 44.04 | 0.0 | 69.1 | 74.5 | None |
Bypassing AI detection is not the same as a faithful rewrite
Undetectable AI recorded the highest bypass rate in the field at 0.957. Nearly 96 percent of the detector tests it faced returned a human verdict, a figure close to the ceiling and well clear of second place. Considered in isolation, that is the highest raw evasion figure of the cycle.
It is also the clearest case this month of an evasion number that does not translate into a good outcome for the person using the tool. Undetectable AI reached that bypass rate by inflating the length of its outputs. Across the cycle it triggered the length-inflation penalty 26 separate times, enough to reach that penalty’s maximum deduction of 10 points against its composite. The rest of its profile is unremarkable. Its meaning preservation (0.730) and its readability (0.559) both sit in the middle of the field, comparable to tools several places below it. The only standout figure in Undetectable AI’s profile was its bypass rate, and it was driven by length inflation rather than by rewrite quality.
The mechanism explains why length inflation is a low-effort tactic rather than a marker of rewrite quality. Commercial detectors estimate how likely a passage was written by a language model, largely from statistical regularities in word choice and sentence construction. Adding length dilutes those signals. A longer passage gives a detector more ordinary, filler-heavy text to average across, which lowers its confidence that any given span is machine-generated. Inflating length is therefore an effective way to move a detector’s score, and of the tactics available to these tools it is the one that demands the least genuine rewriting.
The cost of that route is paid by the user’s text. Someone who pastes a paragraph into a humanizer is asking for that paragraph back: rewritten to read naturally, capable of passing a detector, at roughly the same length and carrying the same meaning. Padding returns something else. It hands back a longer document, thinned out with filler and redundancy, that has to be cut back down before it can be used, and that no longer matches the length the writer intended. A bypass rate produced this way overstates how much the tool actually did for the user. The benchmark’s length penalty exists to measure that overstatement rather than to ignore it. In Undetectable AI’s case, the full 10-point deduction is the gap between the result it reports and the far weaker result its rewrite quality alone would support. A tool that padded less and rewrote better would have finished ahead of it.
Why WriteHuman finished first in July 2026
WriteHuman finished first by combining near-top detector evasion with clean fidelity, rather than by leading any single sub-score. Its bypass rate of 0.816 was the second-highest in the field, behind only Undetectable AI’s padded 0.957 and level with StealthGPT. The difference is in how each tool reached that level. WriteHuman took a single penalty point, for one instance of meaning drift, while the two tools nearest it on bypass each gave back 10 points to fidelity penalties. Its own meaning preservation (0.729) and readability (0.562) are mid-table, so it did not win on rewrite quality either. It finished first because no other tool paired that level of detector evasion with so few distortions to the text.
HumanizerBench is operated by WriteHuman, which is the reason the scoring is mechanical and the underlying data is released in full. The first-place result, like every other result on the board, is recomputed from published files by anyone who runs the verification script, and the weighting that produced it is documented on the methodology page. The purpose of publishing the raw data is to make a claim like “first by balance rather than by any single metric” something a reader can check rather than take on trust.
Length inflation is the most common distortion in the field
Undetectable AI is the extreme case, not the only one. Three tools reached the maximum combined penalty of 10 points this cycle, and they reached it in different ways. Undetectable AI lost all 10 to length inflation alone. ai-humanize-io lost 8 to padding and 2 to meaning drift, and fell four places to tenth as a result. StealthGPT lost across every fidelity category at once: 5 points for padding, 3 for the opposite failure of over-trimming its outputs, and 2 for meaning drift. Below them, Walter Writes gave back 7 points to padding, and Phrasly lost 8 spread across meaning drift and length.
That concentration is not accidental. The market for these tools rewards a single headline number, the bypass rate, and length inflation is the path of least resistance to a higher one. A benchmark that measured evasion alone would place the most aggressive padders at the top of the table and reward exactly the behavior that hurts users most. Measuring fidelity alongside evasion is what prevents that outcome, and this cycle the fidelity penalties reordered most of the board. The cap on each penalty category was raised relative to June, so a padding-heavy strategy costs more than it did in the previous cycle, and the readability sub-score was computed with an updated method this cycle. For those reasons, July’s composites should not be read as directly comparable to June’s on a point-for-point basis.
Meaning drift: when a rewrite changes the original meaning
Length is the most visible fidelity failure because it can be measured directly against the input. Meaning is the subtler one. A rewrite can hold its length and still drift away from what the original text said, and several tools did exactly that. Phrasly recorded the most meaning-drift penalties in the field, losing 4 points to it, despite otherwise respectable readability (0.721). StealthGPT and Stealth Writer each lost 2 points to the same problem.
Meaning drift deserves to be separated from evasion because the two are easily confused. A rewrite that reads as human and clears every detector is still a failure if it no longer means what the writer intended. A tool that evades detection around 60 percent of the time while preserving meaning cleanly can serve a user better than one that evades 80 percent of the time by loosening the text until it barely resembles the source. The penalties for meaning drift are the benchmark’s attempt to hold tools to the second standard as well as the first.
The humanizers that bypass detection without padding
Four tools took no penalties of any kind this cycle: Humanize AI Pro, HIX Bypass, Grammarly, and NoteGPT. The differences among them are instructive. Humanize AI Pro finished third and HIX Bypass sixth, because both combined clean fidelity with genuine detector evasion, at bypass rates of 0.704 and 0.659 respectively. Humanize AI Pro is the clearest illustration of the pattern. It climbed six places from June on a profile with no standout figure: mid-field evasion, solid meaning preservation (0.743), no length manipulation, and no meaning drift. It did not lead a single category. It also lost no points anywhere, and in a cycle shaped by penalties, that was enough for a top-three finish. A competent rewrite that takes no shortcuts scored better than a stronger evasion number bought with padding.
Editing is not evasion: Grammarly and NoteGPT
The other two penalty-free tools describe the opposite situation. Grammarly and NoteGPT both recorded a bypass rate of essentially zero. Not one detector test read either tool’s output as human. Grammarly is not weak at its actual job. It posted the best meaning preservation (0.945) and the best readability (0.821) of any tool in the cycle, both by a clear margin. It simply makes no attempt to disguise text from a detector, because it is an editor rather than a humanizer. It corrects and clarifies writing; it does not try to conceal the writing’s origin. NoteGPT holds the same position with weaker underlying quality.
The contrast matters because the two functions are routinely conflated in marketing and in search results. A tool that improves how writing reads is not the same as a tool that helps writing evade detection, and the July numbers draw that line precisely. The cleanest writer in the field is also the one detectors identified every single time. The distinction matters for anyone selecting a tool, because the two functions serve different purposes, and the highest-quality editor on the board will not move an AI detector at all.
The mid-ranked humanizers: Humbot, Stealth Writer, and Super Humanizer
Below the leaders, the field is tightly grouped. Five tools, Humbot, HIX Bypass, Walter Writes, StealthGPT, and Phrasly, fall within a five-point band between composites of 66.42 and 61.47. In a range that narrow, small penalty differences move a tool several places, which is a further reason to read the sub-scores rather than the rank alone. Humbot, at the top of that band, preserved meaning better than any other tool that seriously attempts evasion, at 0.753, but its outputs earned one of the lowest readability marks in the cycle at 0.439. That combination is a reminder that keeping the meaning intact and reading naturally are separate achievements, and that a tool can manage one without the other. Stealth Writer took fourth on strong evasion (0.812) constrained by lower rewrite quality, including one of the lowest readability scores in the field at 0.445. Super Humanizer sits apart at the low end of the tools that attempt evasion, with a bypass rate of 0.423, well beneath every other humanizer that tries to beat detection.
All 13 tools carry a medium confidence rating, each having been run through the same 33 prompts. No tool was excluded this cycle, which is a change from June, when BypassGPT was removed after failing to return usable output on nearly every prompt. The source passages the tools were asked to rewrite were generated by three separate models, Claude Sonnet 5, Gemini 3.5 Flash, and GPT-5.5, so the results do not reflect the idiosyncrasies of any single text generator.
How to reproduce the July 2026 results
Every figure in this analysis is drawn from the published July cycle. The release
includes the input prompts, every humanized output, every response from all five
commercial detectors, and the scoring script that produced the composites and the
penalties. The full dataset can be cloned from the
public repository, and running
npm run verify recomputes the entire leaderboard from the raw files. The
weighting and the penalty rules are documented on the methodology
page.
The headline result is narrow, but the structure underneath it is consistent. The tools chasing the highest possible bypass rate are, with few exceptions, the same tools that give points back for how they get there. Detection bypass is necessary, and on its own it is not evidence of a good rewrite. Measured against fidelity, the leaders of the July cycle are the tools that cleared detectors without damaging the text they were handed.