The humanizer benchmark that checks your facts
AI humanizers are scored on whether they keep every figure, name and date, add no claims, introduce no errors and leave human writing alone. 151 facts across 36 texts, checked by script.
Results
RemoveAISlop ranks first of the 6 tools that completed every run, with an overall score of 80.1 (range 75.8–82.5).
- RemoveAISlop and Ahrefs AI Humanizer kept every fact in all 30 AI drafts. The rest changed or dropped at least one figure, name or date.
- Clever AI Humanizer passed the AI detector most often (70% of its outputs read as human). For scale, the same detector (ZeroGPT) passed 0% of the untouched AI drafts and 100% of the human-written controls.
- RemoveAISlop publishes this benchmark and ranks first. Its weakest measure is the detector: 3% of its outputs passed, because it changes only the sentences it can rewrite without touching a fact. Judged on the detector alone, Humanize AI (humanizeai.pro) and Clever AI Humanizer beat it. Set your own weights below.
| # | Humanizer | |||||||
|---|---|---|---|---|---|---|---|---|
| 1 | RemoveAISlop freePublisher of this benchmark | 80.1 75.8–82.5 | 100% | 0.03 0.00–0.10 | 0.06 0.00–0.14 | 3% 0%–10% | 1% 0%–3% | 0% |
| 2 | Ahrefs AI Humanizer free, no account | 69.6 65.8–72 | 100% | 0.03 0.00–0.10 | 0.03 0.00–0.08 | 3% 0%–10% | 45% 41%–49% | 20% 13%–25% |
| 3 | Humanize AI (humanizeai.pro) free, no account | 49.5 35–63.8 | 95% 91%–98% | 0.27 0.10–0.50 | 0.17 0.06–0.33 | 50% 33%–67% | 52% 46%–57% | 62% 57%–67% |
| 3= | Scribbr AI Humanizer free, no account (125 words a run) | 37.8 29.4–51.8 | 97% 94%–99% | 0.50 0.20–0.83 | 0.36 0.14–0.61 | 3% 0%–10% | 50% 47%–54% | 0% |
| 3= | Clever AI Humanizer free, no account | 35.9 26–53.5 | 90% 85%–95% | 0.27 0.10–0.47 | 0.50 0.25–0.81 | 70% 53%–87% | 45% 36%–56% | 54% 48%–62% |
| 3= | Undetectable AI free, no account | 8.6 3.9–26.1 | 88% 83%–93% | 0.37 0.10–0.70 | 0.50 0.25–0.81 | 0% | 55% 50%–59% | 35% 27%–45% |
Ranges are 95% bootstrap intervals over the inputs. A rank marked = overlaps the one above it and is a tie.
Still being collected, not ranked
These tools cap free use per day, so their runs are spread over several days. They join the ranking once every input, the human controls and the repeats are in. Until then their numbers cover fewer drafts and their ranges are wider.
| Humanizer | Runs so far | Facts kept | Claims added | Errors added | Detector pass |
|---|---|---|---|---|---|
| StealthWriter free account (10 runs a day) | 10 of 46 | 98% | 0.50 | 1.40 | 60% |
| QuillBot AI Humanizer free account (5 runs a day) | 5 of 46 | 100% | 0.20 | 0.40 | 0% |
| Grammarly AI Humanizer free, no account (about 3 runs a day) | 4 of 46 | 100% | 0.50 | 0.75 | 25% |
Set your own weights
The overall score recomputes as you move them. Measures with no data yet are left out.
What it does that others do not
Facts are checked, not estimated
Each draft is written from a fact sheet, and every fact is looked for in the output exactly. Embedding similarity scores "an average of 6.5 hours" and "around 6 hours" as the same text. We count it as a changed fact.
Human writing is in the test
Six passages written by people go through every tool. A good humanizer leaves them close to how they were, and they show how often each detector flags writing no machine produced.
No AI judges
Every measure is computed by a published script. Errors are counted by LanguageTool, an independent open-source checker, so the publisher is not grading itself.
Registered first, published in full
The method and weights were committed before any result existed. Every prompt, input, output and script is downloadable, and you can set your own weights.
Discover Northwind Analytics, the premier analytics platform built specifically for marketing agencies. Founded in Leeds in 2019, we currently empower 4,200 agencies to streamline their client reporting. Our powerful software saves your team an average of 6.5 hours every single week by automating routine data collection and metric tracking. Seamlessly connect all your campaigns with our native Looker Studio integration for real-time visualization and custom dashboards. Elevate your agency's reporting efficiency, impress clients with crystal-clear insights, and scale your operations effortlessly. Experience the ultimate reporting solution designed exclusively for modern marketing professionals today.
Every highlighted fact has to survive
The draft was written from a fact sheet listing these 6. A humanizer keeps the fact only if it is still there, exactly, or in an equivalent form the sheet lists before the test.
- Northwind Analytics
- Leeds
- 2019
- 4,200 agencies
- 6.5 hours
- Looker Studio
The measures
| Measure | What is computed | Weight |
|---|---|---|
| Facts kept | Share of fact-sheet facts present in the output | 30 |
| Claims added | Promotional words, numbers and names the input did not have | 15 |
| Errors introduced | Spelling and grammar issues the input did not have (LanguageTool) | 15 |
| Detector pass rate | Share of outputs detectors class as human, beside their rate on human text | 20 |
| Human text changed | How much of a human-written passage the tool rewrote | 10 |
| Consistency | How much the output changes between runs of the same input | 10 |
Each measure is scaled across the tools tested and combined with these weights. Every result carries a 95% confidence interval, and tools whose intervals overlap are shown as tied. RemoveAISlop's own AI-phrasing score is shown for reference and never weighted, because it is our catalogue.
Download the data
- METHOD.mdMethod, weights and changelog
- fact-sheets.jsonFact sheets: the ground truth
- prompts.jsonThe exact prompt for every draft
- input-texts.jsonEvery input, as each tool received it
- inputs.jsonInputs with their facts and sources
- generation-log.jsonHow each draft was generated
- repeat-set.jsonInputs run three times for consistency
- outputs.jsonEvery output, from every tool
- leaderboard.jsonEvery score with its confidence interval
- detectors.jsonEvery detector verdict, with its score
- tools.jsonThe tools and the plan each ran on
- collection-log.jsonHow each tool was run, and what went wrong
Data CC BY 4.0, scripts MIT. Raw outputs are added when the cycle publishes.
Cite this benchmark
RemoveAISlop (2026). The Humanizer Benchmark, November 2026 cycle. https://removeaislop.com/humanizer-benchmark
@misc{humanizerbenchmark202611,
title = {The Humanizer Benchmark: {November 2026} cycle},
author = {{RemoveAISlop}},
year = {2026},
url = {https://removeaislop.com/humanizer-benchmark},
note = {Open data and scoring scripts}
}We make one of the tools tested.
RemoveAISlop runs this benchmark and is tested in it, through the same free humanizer anyone can use. So nothing here depends on trusting us: the method came first, the scoring is a script, every output is public, and the weights are yours to change. Results where we lose are shown the same way as results where we win.
Questions people ask
Try the humanizer we test
Free up to 1,200 words a draft. Rules, not a language model.