Test log
Humanize AI Text test data
Raw results for the Humanize AI Text review. The verdict is in the review.
- Tested
- 6 November 2026
- Tester
- Miles Foster
- Account
- Free demo account, signed up by RankDude, nothing paid to the brand
- Tests
- 8
- Screenshots
- 7
- Method
- Published methodology
- Environment
- Google Chrome on desktop
Results
| ID | Test | Key measurement | Score |
|---|---|---|---|
| T01 | Detection-Bypass Rate | Rewrites read as human by both detectors: 0 of 12 | 1/5 |
| T02 | Coverage Across Detectors | Rewrites read as human by ZeroGPT: 6 of 11 | 2/5 |
| T03 | Meaning & Fact Preservation | Rewrites keeping every fact and the meaning: 5 of 8 | 2/5 |
| T04 | Readability & Grammar | Rewrites with an error: 10 of 12 | 1/5 |
| T05 | Non-English Support | Passages that passed and kept their meaning: 0 of 4 | 1/5 |
| T06 | Input, Length & Output Support | Points that worked: 2 of 10 | 1/5 |
| T07 | Humanizing Speed | Median time: 3.0 s | 5/5 |
| T08 | Pricing & Free-Tier Limits | Free words per run: about 880 | 5/5 |
| Overall (in the review) | 2.5/5 | ||
Detection-Bypass Rate
1/5Tested- Method
- Twelve AI passages were humanized once each and the first rewrite kept.
- Rewrites read as human by both detectors
- 0 of 12
- Rewrites read as human by ZeroGPT
- 6 of 11
- Rewrites read as human by Sapling
- 1 of 12
- Remark
- Scale 1: fewer than 15 of 50 bypassed. 0 of 12 is 0 percent, scaled to the sample. Only two of the four detectors in the method could be reached.
Raw notes
Twelve AI passages were humanized once each and the first rewrite kept. Each original and each rewrite was scanned once in ZeroGPT and in Sapling. A rewrite counts as bypassed only when every detector reads it as human (50 percent AI or below).
No rewrite passed both detectors. ZeroGPT read 6 of 11 rewrites as human; Sapling read 1 of 12.
ZeroGPT counted the Chinese text as 0 words and asked for more text, so that pair was not counted for ZeroGPT. All twelve originals read as AI in Sapling, and eleven in ZeroGPT.
- Twelve passages and two detectors, not 50 and four. A free tool with no account cannot be tested against Originality.ai or Turnitin.
Coverage Across Detectors
2/5Tested- Method
- The same scans were split by detector, and the grade is set by the detector the tool beats least often.
- Rewrites read as human by ZeroGPT
- 6 of 11
- Rewrites read as human by Sapling
- 1 of 12
- Remark
- Scale 2: one detector beaten under half the time. Sapling is that detector. Turnitin and Originality.ai, the two a student or publisher is checked by, could not be tested.
Raw notes
The same scans were split by detector, and the grade is set by the detector the tool beats least often.
ZeroGPT: 6 of 11 rewrites read as human (55 percent). The passes were the remote work, exercise, Industrial Revolution, electric cars and social media passages and the French one.
Sapling: 1 of 12 (8 percent). Only the climate change rewrite got under 50 percent, at 41.5. Every other rewrite scored 95 percent AI or more.
- Two detectors only. The page's own row of ticks was not counted.
Meaning & Fact Preservation
2/5Tested- Method
- The eight English passages carry names, dates, figures and a stated reason written in before the tool was opened.
- Rewrites keeping every fact and the meaning
- 5 of 8
- Rewrites keeping every name, date and figure
- 8 of 8
- Remark
- Scale 2: 14 to 19 of 30 clean. 5 of 8 is scaled to about 19 of 30. An invented or softened claim counts as a failure, as the method says.
Raw notes
The eight English passages carry names, dates, figures and a stated reason written in before the tool was opened. Each rewrite was read against its list.
Every name, date and figure survived in all eight rewrites, for example Hargreaves in 1764, Watt in 1776 and the Factory Act of 1833.
Five rewrites kept the meaning cleanly. Three did not: the remote work rewrite adds "a smoother routine", the exercise rewrite softens the endorphin claim to "or at least help you feel better", and the social media rewrite invents "people start doubting everything, even when it's not deserved".
- Eight passages, not 30.
Readability & Grammar
1/5Tested- Method
- All twelve rewrites were read in full and marked clean or broken.
- Rewrites with an error
- 10 of 12
- Remark
- Scale 1: more than 22 of 50 broken. 10 of 12 is 83 percent, scaled to the sample.
Raw notes
All twelve rewrites were read in full and marked clean or broken. A broken passage has a grammar or spelling error the original did not, or a stray character.
Ten of twelve rewrites are broken. A space before a comma appears in five ("in 1776 ,"), double spaces in two, and the Spanish rewrite dropped the marks over its letters ("mas relevantes").
Grammar slips include "In this today's ever shifting digital scene" and "less moving components". Filler such as "kinda", "yeah" and "honestly" is added throughout. Only the time management and social media rewrites read clean.
- Twelve rewrites, not 50. No readability score was run.
Non-English Support
1/5Tested- Method
- One AI passage each in Spanish, French, German and Chinese was humanized once, scanned in both detectors and read for meaning and grammar.
- Passages that passed and kept their meaning
- 0 of 4
- Remark
- Scale 1: 2 or fewer of 12 pass. 0 of 4 is 0 percent. There were no human controls in other languages.
Raw notes
One AI passage each in Spanish, French, German and Chinese was humanized once, scanned in both detectors and read for meaning and grammar.
None of the four passed both detectors and kept clean grammar. ZeroGPT passed the French rewrite (0% AI) but Sapling read it as 99.1% AI. The Spanish and German rewrites scored 62.6% and 59.5% in ZeroGPT.
Facts survived in all four. Grammar did not: Spanish lost letter marks, French wrote "une espèce de habitude", and the Chinese rewrite has stray spaces inside words and an odd phrase, "对号入座", that the original did not have.
- Four passages, not twelve, and no human controls.
Input, Length & Output Support
1/5Tested- Method
- Each of the ten points was tried on the page on the test date, with 2,600, 2,000, 1,500, 1,000 and 700 words pasted to find the cap.
- Points that worked
- 2 of 10
- Cap per run
- 6,000 characters
- Remark
- Scale 1: 3 or fewer points, paste with a tight cap. The Expand button lengthened a rewrite to 192 words but is not a strength choice.
Raw notes
Each of the ten points was tried on the page on the test date, with 2,600, 2,000, 1,500, 1,000 and 700 words pasted to find the cap.
Worked: pasting text, and the built-in detector row under the output. That is 2 of 10.
Not available: a file upload (the page has no file box), a batch, a choice of strength or mode, a word cap of 2,000 words (the cap is 6,000 characters, about 880 words), a browser extension and an API. Not tested: formatting through the rewrite and the Copy button.
- Two points were not tried: the Copy button and formatting through the rewrite.
Humanizing Speed
5/5Tested- Method
- All twelve rewrites were timed from the click on Humanize until the output stopped changing, checking four times a second.
- Median time
- 3.0 s
- Slowest rewrite
- 3.3 s
- Remark
- Scale 5: a median under 8 seconds. One second of waiting for the text to stop changing was taken off each time.
Raw notes
All twelve rewrites were timed from the click on Humanize until the output stopped changing, checking four times a second.
Median 3.0 seconds, fastest 2.7, slowest 3.3. No rewrite failed or queued.
- Passages of 86 to 157 words. A full 880-word run was not timed.
Pricing & Free-Tier Limits
5/5Tested- Method
- The page was used until it stopped me, and its home, about, contact and terms pages and the Premium link were read on the test date.
- Free words per run
- about 880
- Account needed
- No
- Remark
- Scale 5: a free tier of 500 words a day or more.
Raw notes
The page was used until it stopped me, and its home, about, contact and terms pages and the Premium link were read on the test date.
The tool is free with no account and no card, and the page says usage is unlimited. The only limit I met is 6,000 characters a run, about 880 words, which is more than 500 words a day.
The Try Premium link leads to a different site, humanizeai.io. Its Standard plan is $20 a month, or $4 a month billed yearly, with Premium at $30 and Elite at $40. I did not buy anything there or run its modes. The terms say that by using the service you agree to receive newsletters and promotions, with an opt-out.
- Whether the free tier runs the same engine as Premium was not checked.
Coverage
Tested
- 12 rewrites (8 English, 1 each Spanish, French, German and Chinese) timed from the click to the finished text
- 24 detector scans in ZeroGPT and 24 in Sapling: every original and every rewrite, once each
- The per-run size limit, found by pasting 2,600, 2,000, 1,500, 1,000 and 700 words
- The Expand, Use Again and Compare buttons and the detector row under the output
- The about, contact and terms pages and the Premium plans page read on the test date
Not tested or limited
- Twelve passages, not the 50 of the methodology, and all written by one model (Claude) rather than six
- Two detectors, not four. GPTZero's free box did nothing without a login, Copyleaks did not take pasted text in the held browser, and Originality.ai and Turnitin need an account
- Plagiarism introduced was not run (no checker without an account) and natural voice was not run (no blind reading against human controls)
- En1, en2 and en3 were humanized twice because the first screenshots were badly cropped; the later rewrite is the one scanned and judged
- The Copy button, headings, lists, bullet points and pasted formatting were not tested
- Premium on humanizeai.io is a separate product; no plan was bought and its modes were not run
- Sapling's scan of the rewritten Industrial Revolution passage was repeated once for the screen recording; it gave the same 99.2%, and the first scan is the one counted
- ZeroGPT did not return a verdict for the Spanish and German originals on the first scan (the wait was too short); those two originals were scanned again and the second verdict counted
Still to test
- The other 38 passages and four more models
- GPTZero, Originality.ai, Copyleaks and Turnitin on an account
- Plagiarism introduced, with a plagiarism checker
- Natural voice, read blind against human controls
- Premium on humanizeai.io and its Aggressive and Enhanced modes
- Headings, lists and the Copy button