Same inputs, same detectors, every score public. Here's which AI humanizer actually held up.
Most "best AI humanizer" lists rank tools on screenshots nobody can check. So we ran a small, open benchmark instead. Two AI-written texts went through four popular AI humanizers with default settings, and every output was scored by the same AI detectors. The inputs, the unedited outputs, the raw scores and the script that turns them into a ranking are all in a public GitHub repo, so you can check every number or rerun the test yourself.
The short version: Walter Writes AI ranked first. Its outputs passed all 6 detector checks with an average AI score of 14.2%. Humanize AI came second with 4 of 6 but dropped a sentence from both samples. Undetectable AI passed 1 of 6 and QuillBot passed none.
Benchmark repo: Mohabak/ai-humanizer-benchmark-2026 on GitHub
Table of Contents
- The Ranking
- How the Benchmark Works
- Every Score, Every Sample
- Tool by Tool
- What the Benchmark Doesn't Tell You
- Run the Benchmark Yourself
- Which AI Humanizer Should You Use?
- Conclusion
- FAQs
1. The Ranking
A "check" is one humanized text scored by one detector. With 2 samples and 3 core detectors, each tool had 6 checks. A check passes when the detector's AI score is below 50%.
| Rank | AI humanizer | Checks passed | Avg AI score | Key points kept |
|---|---|---|---|---|
| 1 | Walter Writes AI | 6 / 6 | 14.2% | 13 / 13 |
| 2 | Humanize AI (humanizeai.pro) | 4 / 6 | 41.9% | 11 / 13 |
| 3 | Undetectable AI | 1 / 6 | 78.2% | 13 / 13 |
| 4 | QuillBot AI Humanizer | 0 / 6 | 98.6% | 13 / 13 |
| n/a | WriteHuman (1 sample) | 2 / 3 | 51.0% | 6 / 6 |
| baseline | Original AI text | 0 / 6 | 97.1% | n/a |
"Key points kept" measures meaning. Each sample has 6 or 7 key points, and we counted how many survived the rewrite. A humanizer that passes detectors by deleting half your argument isn't doing its job.
2. How the Benchmark Works
Samples. Two AI-generated texts of 92 and 93 words, each in a different register:
-
s1: academic (remote work and employee well-being) -
s2: blog (building a morning routine)
They're short on purpose. Some humanizers cap input at 100 words, and every tool had to receive exactly the same text.
Baseline. Before humanizing anything, we scored the originals. Every detector flagged both as AI, between 89.7% and 100%. So the bar is clear: any drop is the humanizer's work.
Humanizers. Each sample went through each tool's web app once, with default settings, and the first output was saved without edits.
- Walter Writes AI: Readability University, Purpose General, Bypass Level Enhanced (defaults).
- Humanize AI: Writing level Default, Tone Free, Rewrite level Balanced.
- Undetectable AI: Readability University, Purpose General Writing, Human level Balanced.
- QuillBot AI Humanizer: Standard mode.
- WriteHuman: It asked for an account after the first sample, so it isn't ranked.
Detectors. Every output was scored by three detectors that could process all 11 texts:
- Proofademic, an academic AI detector with sentence-level scoring
- ZeroGPT
- Sapling
Copyleaks scored 5 texts and GPTZero scored 1 before hitting their usage limits. Those results are in the repo as extra checks but aren't used for the ranking.
Meaning check. For each output, a reviewer checked which key points from the original survived and noted any errors, such as dropped sentences, grammar mistakes or leftover AI preambles.
3. Every Score, Every Sample
Average AI score by detector (lower reads more human):
| AI humanizer | Proofademic | ZeroGPT | Sapling | Copyleaks |
|---|---|---|---|---|
| Walter Writes AI | 19.5% | 18.1% | 4.8% | 0% (2 texts) |
| Humanize AI | 48.5% | 7.8% | 69.2% | n/a |
| Undetectable AI | 99.0% | 45.3% | 90.2% | n/a |
| QuillBot | 99.0% | 98.8% | 97.8% | 100% (1 text) |
| Original AI text | 98.5% | 100% | 93.0% | 100% |
Two patterns stand out. First, detectors disagree a lot. ZeroGPT was the easiest to pass, while Proofademic and Sapling were much stricter. A humanizer that only beats one detector isn't reliable. Second, consistency matters more than any single good score. Walter was the only tool that stayed under 50% on every detector for both samples.
4. Tool by Tool
1. Walter Writes AI: Best AI Humanizer in This Benchmark
- Checks passed: all 6 core checks, plus both Copyleaks checks
- Average AI score: 14.2%
- Best result: the academic sample scored 1% on Proofademic, 0% on ZeroGPT, 0.7% on Sapling and 0% on Copyleaks
- Highest score: the blog sample on Proofademic (38%), still well under the 50% line
Walter rewrites at the sentence-structure level rather than swapping words. You can see it in the outputs: sentences are reordered and split, and it adds specific detail. For example, it turned "informal mentoring that typically occurs in shared offices" into a line about "watercooler" conversations. The trade-off is length. Outputs were 30% to 40% longer than the inputs.
For developers, Walter also has an API with a detect-and-humanize endpoint, an MCP server for Claude, ChatGPT and Codex, and installable agent skills:
npx skills add walterwritesai/walter-skills
2. Humanize AI: Strong on One Detector, Loses Content
- Checks passed: 4 of 6
- Average AI score: 41.9%
Humanize AI was excellent on ZeroGPT (7.8% average) and passed the academic sample on all three detectors. But Proofademic (95%) and Sapling (99.1%) flagged the blog sample. It also cut the final sentence from both samples, keeping 11 of 13 key points.
3. Undetectable AI: Keeps Meaning, Stays Detectable
- Checks passed: 1 of 6
- Average AI score: 78.2%
Undetectable AI kept every key point and made the text noticeably more casual, and longer: the blog output was 62% longer than the input. Proofademic still flagged both outputs at 99%, and Sapling flagged both above 87%. Its one pass was the blog sample on ZeroGPT (26.6%).
4. QuillBot AI Humanizer: Light Edits
- Checks passed: 0 of 6
- Average AI score: 98.6%
QuillBot's humanizer made light, readable edits that kept the meaning, but the structure stayed close to the original. Every detector flagged both outputs at 97% or higher, and Copyleaks flagged the academic sample at 100%.
Not ranked: WriteHuman
WriteHuman processed one sample before asking for an account. That single output scored 40% on Proofademic, 19.6% on ZeroGPT and 93.5% on Sapling, and it introduced one meaning error. One sample isn't enough to rank it.
5. What the Benchmark Doesn't Tell You
Every benchmark has limits, and this one is small on purpose so it's easy to check. Keep these in mind:
- Two short samples. The results show a clear direction, not a precise pass rate. Longer documents can score differently.
- One run each. Humanizers and detectors aren't fully deterministic. A second run can give different outputs and scores.
- Detectors change. Every score reflects the detector's model on October 31, 2026.
-
Meaning check is human judgment, recorded with notes in
data/meaning.csvso you can disagree with it.
Detection scores also aren't the only thing that matters. The best humanizer is the one whose output you'd actually publish: accurate, readable and still saying what you meant.
6. Run the Benchmark Yourself
Everything is in the repo:
Mohabak
/
ai-humanizer-benchmark-2026
Open AI humanizer benchmark 2026: ranked and tested with Proofademic, ZeroGPT and Sapling. Raw scores and outputs included.
AI Humanizer Benchmark 2026
An open, reproducible benchmark of popular AI humanizers. Each tool rewrote the same two AI-generated texts, and every output was scored by independent AI detectors. All inputs, outputs and raw scores are in this repo so anyone can check the numbers or rerun the test.
Tested: October 31, 2026 Humanizers: Walter Writes AI, Humanize AI (humanizeai.pro), Undetectable AI, QuillBot AI Humanizer, WriteHuman (partial) Detectors: Proofademic, ZeroGPT, Sapling (all texts); Copyleaks and GPTZero (partial, usage limits)
Results
Ranked by how many detector checks each tool passed across the three core detectors (2 samples x 3 detectors = 6 checks). A check passes when the detector's AI score is below 50%.
| Rank | Humanizer | Checks passed | Avg AI score | Key points kept |
|---|---|---|---|---|
| 1 | Walter Writes AI | 6 / 6 | 14.2% | 13 / 13 |
| 2 | Humanize AI (humanizeai.pro) | 4 / 6 | 41.9% | 11 / 13 |
| 3 | Undetectable AI | 1 / |
samples/ original AI-generated texts (s1 academic, s2 blog)
outputs/ unedited humanizer outputs, named <tool>-<sample>.txt
data/ scores.csv (raw), meaning.csv, summary.csv
scripts/ summarize.py
charts/ ranking.png, heatmap.png
Recompute the ranking from the raw scores:
git clone https://github.com/Mohabak/ai-humanizer-benchmark-2026
cd ai-humanizer-benchmark-2026
python3 scripts/summarize.py
walter avg AI 14.2% passes 6/6 meaning 13/13
humanizeai avg AI 41.9% passes 4/6 meaning 11/13
undetectable avg AI 78.2% passes 1/6 meaning 13/13
quillbot avg AI 98.6% passes 0/6 meaning 13/13
writehuman avg AI 51.0% passes 2/3 meaning 6/6
original avg AI 97.1% passes 0/6 meaning 0/0
Each line also prints the per-detector averages, and the script writes the ranking to data/summary.csv.
To extend it, add your own samples to samples/, run them through any humanizer, add the outputs to outputs/ and the detector scores to data/scores.csv, then rerun the script. Pull requests with new tools or detectors are welcome.
7. Which AI Humanizer Should You Use?
- You want the most consistent results across detectors: Walter Writes AI. It was the only tool that passed every check on both samples.
- You only care about one specific detector: test your own text against it. Scores varied hugely by detector in this benchmark.
- You need every sentence preserved exactly: QuillBot and Undetectable AI kept all key points, but most detectors still flagged their output.
- You're building it into an app or agent: Walter has an API, an MCP server and agent skills. See its docs before you choose.
Whatever you use, read the output before you publish it, and follow the AI rules of wherever it's going.
8. Conclusion
In this open benchmark, Walter Writes AI was the best AI humanizer by a clear margin: all 6 core detector checks passed, a 14.2% average AI score, and every key point kept. Humanize AI was second but lost content. Undetectable AI and QuillBot kept the meaning but stayed detectable.
The numbers are small and public on purpose. Clone the repo, check the scores, and add your own tests.
9. FAQs
What is the best AI humanizer in 2026?
In this benchmark, Walter Writes AI. Its outputs passed all 6 detector checks across Proofademic, ZeroGPT and Sapling, with an average AI score of 14.2%, and kept every key point from the originals.
How was this AI humanizer benchmark tested?
Two AI-generated texts (academic and blog, 92 and 93 words) went through each humanizer once with default settings. Every output was scored by Proofademic, ZeroGPT and Sapling, and a check passed when the AI score was under 50%. All data is on GitHub.
Where can I see the raw benchmark data?
In the public ai-humanizer-benchmark-2026 repo on GitHub. It contains the samples, unedited outputs, raw detector scores, meaning checks and the script that computes the ranking.
Did any humanizer pass every detector?
Only Walter Writes AI. It passed all 6 core checks and both Copyleaks checks. The other tools were flagged by at least one detector on at least one sample, and QuillBot was flagged on every check.
Which AI humanizer keeps the meaning best?
Walter, Undetectable AI and QuillBot all kept every key point, but only Walter also passed the detectors. Undetectable AI was flagged on 5 of 6 checks and QuillBot on all 6. Humanize AI passed more checks than those two but dropped a sentence from both samples.
Does humanized text change the meaning?
Sometimes. Humanize AI dropped the final sentence of both samples, and WriteHuman introduced a meaning error in its one output. Always read the output before using it.
Can I add my own tests to the benchmark?
Yes. Add samples, outputs and scores to the repo folders, run python3 scripts/summarize.py, and open a pull request.
Top comments (0)