An AI humanizer rewrites AI-generated text so that detectors read it as human. Two things matter: does the rewrite get past detectors, and does it still say what you wrote? We tested both on the same outputs. This is our own test, run in October 2026, on free tiers only.
How we tested
- Five passages written with AI on Oct 9, 145 to 167 words each: a student essay, a product description, a news story, a science abstract and a cover letter. Each one plants 8 facts we can check (5 numbers or dates, 3 names), 2 hedges ("about", "may"), 1 negation and 1 quote. The people and figures are invented, so a tool can't "correct" them from what it already knows.
- 11 humanizers, free tier, default settings. Six ran with no login on Oct 9 (Grammarly, HIX Bypass, Humanize AI Pro, HumanizeAIText, Undetectable AI's free Basic model, WriteHuman). Five ran with a free account on Oct 23 (GPTHuman, Netus AI, Phrasly, QuillBot, Ryne AI). Free limits meant most of the account tools could only run the student essay. QuillBot's free plan takes 100 words at a time, so its passages ran in chunks.
- Five detectors read the 23 outputs and the 5 originals on Oct 23: GPTZero and Pangram with free accounts, Sapling and ZeroGPT without (QuillBot's detector read the first-round outputs on Oct 23). One run per text. Pangram's free credits covered 10 texts, so it read the student essay from every tool it could.
- We read every output against its source and counted the facts kept, changed, dropped or invented, the hedges kept, the negation and the quote.
The same essay, through every tool
The student essay (P1) is the one passage every humanizer ran. Detector columns show the share each tool read as AI; "–" means not checked.
| Humanizer | GPTZero | Pangram | Sapling | ZeroGPT | Facts kept (of 8) | Claims added | Quote |
|---|---|---|---|---|---|---|---|
| GPTHuman | 100% | 100% | 0.1% | 0% | 8 | 0 | verbatim |
| Grammarly | 100% | 100% | 0.2% | 0% | 8 | 0 | verbatim |
| HIX Bypass | 2% | 100% | 90.4% | 0% | 5 | 1 | unquoted |
| Humanize AI Pro | 5% | 2% | 99.7% | 0% | 8 | 1 | verbatim |
| HumanizeAIText | 100% | – | 1.2% | 0% | 7 | 1 | altered |
| Netus AI | 100% | 100% | 0% | 0% | 7 | 0 | altered |
| Phrasly | 0% | 0% | 0.1% | 11.4% | 8 | 1 | verbatim |
| QuillBot | 100% | 100% | 46.2% | 17.8% | 8 | 0 | verbatim |
| Ryne AI | 100% | 100% | 0% | 0% | 8 | 0 | altered |
| Undetectable AI | 100% | – | 99.6% | 0% | 8 | 0 | altered |
| WriteHuman | 0% | 35% (AI Detected) | 0% | 0% | 8 | 2 | verbatim |
No essay was read as human by every detector and also kept its meaning intact. Phrasly's essay scored low on all four detectors that read it, but it added a claim the source never made. Grammarly's essay was the only one that kept every fact, hedge and the quote word for word, and GPTZero and Pangram both read it as 100% AI.
What the detectors did
With the five AI originals: GPTZero read all five as 100% AI. Sapling read four of five as AI and scored the science abstract 0%. ZeroGPT gave them 19% to 73%. QuillBot's detector scored all five 0%. Pangram read the one original it checked as 100% AI.
| Original | GPTZero | Pangram | Sapling | ZeroGPT | QuillBot |
|---|---|---|---|---|---|
| Student essay | 100% | 100% | 95.4% | 73% | 0% |
| Product description | 100% | – | 99.6% | 61.7% | 0% |
| News story | 100% | – | 98.1% | 55.8% | 0% |
| Science abstract | 100% | – | 0% | 56.8% | 0% |
| Cover letter | 100% | – | 93.1% | 19.3% | 0% |
With the 23 humanized outputs: GPTZero read 16 of 23 as mostly AI, 10 of them with its "possible AI paraphrasing" label. Pangram flagged 7 of the 9 it checked. Sapling read 8 of 23 as mostly AI, and ZeroGPT 0 of 23.
The detectors often disagreed on the same text. HIX Bypass's essay scored 2% on GPTZero and 100% on Pangram. Humanize AI Pro's essay scored 5% on GPTZero and 2% on Pangram, but 99.7% on Sapling. The essays from GPTHuman, Netus AI and Ryne AI scored 100% on GPTZero and Pangram and 0% to 0.1% on Sapling. We don't know why they split this way: each detector reports a score, not its reasons.
What the humanizers did to the text
Across all 23 outputs, 173 of 184 planted facts survived. 8 outputs changed or dropped at least one fact, 13 added a claim that isn't in the source, 14 of 46 hedges were removed, and the quote was no longer word for word in 14. Only 2 outputs came back clean: Grammarly's essay and QuillBot's product description.
Some of the changes, quoted from the outputs:
- HIX Bypass turned "about 100 million acres" into "about 40 million acres", and "tenant farmers" into "tenant miners".
- Netus AI turned "fallen by 40 percent" into "dropped by 40", and in the product description made the trail guide "she" and the side pockets "two side pockets".
- Phrasly added "With a century of rainfall data available, farmers knew there was risk of drought", which the source doesn't say.
- QuillBot turned "Relief programs did not reach most tenant farmers" into "Most tenant farmers did not participate in relief programs".
- HumanizeAIText renamed the historian Margaret Ellison "Margaret Eller".
- Ryne AI turned "Some economists may disagree" into "Certainly, some economists would disagree" and dropped "about" from "about 100 million acres".
If you use a humanizer on anything that will be graded, cited or published, read the output against your draft line by line. Numbers, names and hedges are where the changes hide.
Every output
| Humanizer | Passage | GPTZero | Pangram | Sapling | ZeroGPT | Facts kept | Claims added | Hedges removed | Quote |
|---|---|---|---|---|---|---|---|---|---|
| GPTHuman | student essay | 100% | 100% | 0.1% | 0% | 8/8 | 0 | 1/2 | verbatim |
| Grammarly | student essay | 100% | 100% | 0.2% | 0% | 8/8 | 0 | 0/2 | verbatim |
| HIX Bypass | student essay | 2% | 100% | 90.4% | 0% | 5/8 | 1 | 0/2 | unquoted |
| Humanize AI Pro | student essay | 5% | 2% | 99.7% | 0% | 8/8 | 1 | 1/2 | verbatim |
| HumanizeAIText | student essay | 100% | – | 1.2% | 0% | 7/8 | 1 | 0/2 | altered |
| Netus AI | student essay | 100% | 100% | 0% | 0% | 7/8 | 0 | 1/2 | altered |
| Phrasly | student essay | 0% | 0% | 0.1% | 11.4% | 8/8 | 1 | 0/2 | verbatim |
| QuillBot | student essay | 100% | 100% | 46.2% | 17.8% | 8/8 | 0 | 2/2 | verbatim |
| Ryne AI | student essay | 100% | 100% | 0% | 0% | 8/8 | 0 | 2/2 | altered |
| Undetectable AI | student essay | 100% | – | 99.6% | 0% | 8/8 | 0 | 1/2 | altered |
| WriteHuman | student essay | 0% | 35% | 0% | 0% | 8/8 | 2 | 0/2 | verbatim |
| Humanize AI Pro | product description | 100% | – | 3.7% | 0% | 8/8 | 2 | 0/2 | verbatim |
| HumanizeAIText | product description | 53% | – | 0% | 0% | 7/8 | 1 | 0/2 | rewritten |
| Netus AI | product description | 100% | – | 0% | 0% | 8/8 | 2 | 0/2 | dropped |
| QuillBot | product description | 100% | – | 95.3% | 0% | 8/8 | 0 | 0/2 | verbatim |
| Undetectable AI | product description | 100% | – | 99.4% | 0% | 6/8 | 5 | 2/2 | dropped |
| Humanize AI Pro | news story | 0% | – | 0.1% | 0% | 8/8 | 1 | 0/2 | rewritten |
| HumanizeAIText | news story | 100% | – | 85.5% | 0% | 8/8 | 0 | 0/2 | altered |
| Undetectable AI | news story | 100% | – | 99.6% | 0% | 7/8 | 1 | 0/2 | altered |
| Humanize AI Pro | science abstract | 0% | – | 0% | 10% | 7/8 | 0 | 1/2 | verbatim |
| HumanizeAIText | science abstract | 38% | – | 0% | 0% | 8/8 | 0 | 0/2 | rewritten |
| Humanize AI Pro | cover letter | 100% | – | 89.8% | 0% | 8/8 | 1 | 1/2 | rewritten |
| HumanizeAIText | cover letter | 100% | – | 0% | 0% | 7/8 | 3 | 2/2 | dropped |
Limits of this test
- It is small: 23 outputs, one run each. Humanizers and detectors don't give the same result twice, so a single score can move.
- Free tiers and default settings only. Paid modes may behave differently.
- The passages are short (145 to 220 words after rewriting), and every detector says short text is harder to judge.
- The originals were written with one AI model; detectors may score other models' text differently.
- Not run: Twixify (server errors on Oct 23) and Walter Writes (its free trial had ended). Originality.ai and Copyleaks were left out because their terms limit publishing results from their detectors.
- There is no Turnitin result here. Turnitin sells only to schools, so we can't test it; see does Turnitin detect humanizers? for what Turnitin says itself.
Every reading is in the CSV: one row per output, with each detector's score and our fact counts. For how often detectors flag human writing, see AI detector false positives; for one detector in depth, is GPTZero accurate?