AI humanizer: what it does, and how to judge one
An AI humanizer takes text a detector reads as machine-written and rewrites it so it reads as a person's. Most fail in one of two ways: they don't move the score, or they move it by changing what you said.
Updated October 1, 2026
What a humanizer has to change
Detectors don't keep a list of banned words. They're models that read the whole texture of a piece: how predictable each word is given the ones before it, how even the sentences are, how ideas are joined. A humanizer that only swaps vocabulary leaves all of that in place.
We tested the vocabulary approach first. On 300 AI drafts, our rules removed stock phrases, filler and chatbot openers and swapped words like "utilize" for plain ones. The detector still flagged 92.7% of them, against 93.3% before (pilot set, MELD at its 1% line).
What a sentence-level rewrite looks like
A real pair from our round-3 test: an AI draft of a LinkedIn post, written by Qwen3.5, and the rewrite by our model. The detector scored the draft 5.27 (flagged above 4.10) and the rewrite -1.75. Meaning similarity to the draft: 0.95 out of 1.
- Antarctica holds 90% of the world's ice and 70% of its fresh water, yet these vital resources face unprecedented threats from warming air and waters.
- RewriteAntarctica has 90% of the world's ice and 70% of its fresh water, but warming air and waters pose unprecedented threats.
- Protecting the Southern Ocean is a defining challenge of our time, offering a unique opportunity to unite nations for a common cause.
- Rewrite(cut)
- Please join us immediately at only.one/antarctica. Every signature counts. […] Let's act decisively to secure our planet's future.
- RewritePlease sign the petition immediately at only.one/antarctica Antarctica needs us more than ever!
| AI draft (excerpts) | Rewrite |
|---|---|
| Antarctica holds 90% of the world's ice and 70% of its fresh water, yet these vital resources face unprecedented threats from warming air and waters. | Antarctica has 90% of the world's ice and 70% of its fresh water, but warming air and waters pose unprecedented threats. |
| Protecting the Southern Ocean is a defining challenge of our time, offering a unique opportunity to unite nations for a common cause. | (cut) |
| Please join us immediately at only.one/antarctica. Every signature counts. […] Let's act decisively to secure our planet's future. | Please sign the petition immediately at only.one/antarctica Antarctica needs us more than ever! |
Notice what changed: the stock framing ("a defining challenge of our time", "every signature counts") went, the facts and numbers stayed, and the rhythm got rougher. Notice too that a whole sentence was dropped. A similarity score of 0.95 doesn't mean nothing was lost, which is why you read the result before using it.
How 20 commercial humanizers did against 7 detectors
Two public benchmarks publish about 2,400 rewrites by about 20 commercial humanizers, each with verdicts from GPTZero, Originality, Copyleaks, Winston, ZeroGPT and others. We added MELD's verdict to every rewrite.
- The spread is huge. The share of a tool's rewrites that GPTZero let through as human ranged from 0% to 88% across tools.
- Detectors disagree. Of the rewrites MELD passed, GPTZero still flagged 33–36%, Originality 22–33% and Copyleaks 16–25%. Rank agreement between MELD and the others was 0.3 to 0.6, where 1 would mean they order texts the same way.
- One tool passed MELD 68–71% of the time but GPTZero only 11–22%. Tuning against one detector can leave you exposed to another.
What to test before you trust one
- Which detector was it tested on?
- Why it mattersPassing one detector says little about another, as the numbers above show. A tool that won't name its detector hasn't shown you anything.
- How often does that detector flag real human text?
- Why it mattersIf the detector flags 10% of human writing, "reads as human" is a low bar, and a 5% result may be noise.
- One try, or the best of several?
- Why it mattersPicking the best of four with the same detector that grades the result inflates the number. The one-try rate is the fair one.
- Does it keep your meaning?
- Why it mattersThe easiest way to fool a detector is to write something else. Compare the facts, names and numbers in the output against your draft.
- Does it add things?
- Why it mattersRewriters trained on the web sometimes restore a figure or a title they remember. Ours once added a job title the draft didn't have.
| Check | Why it matters |
|---|---|
| Which detector was it tested on? | Passing one detector says little about another, as the numbers above show. A tool that won't name its detector hasn't shown you anything. |
| How often does that detector flag real human text? | If the detector flags 10% of human writing, "reads as human" is a low bar, and a 5% result may be noise. |
| One try, or the best of several? | Picking the best of four with the same detector that grades the result inflates the number. The one-try rate is the fair one. |
| Does it keep your meaning? | The easiest way to fool a detector is to write something else. Compare the facts, names and numbers in the output against your draft. |
| Does it add things? | Rewriters trained on the web sometimes restore a figure or a title they remember. Ours once added a job title the draft didn't have. |
- 01Take three of your own AI drafts of the kind you actually write.
- 02Run each through the humanizer once. Don't regenerate until it passes.
- 03Check each result with a detector the humanizer doesn't advertise, and note the score before and after.
- 04Read every output against its draft, line by line, for changed facts, added names and dropped points.
- 05Run one piece of your own human writing through the same detector, so you know its false-alarm rate on your style.
How Patois scored
Patois is an open 9-billion-parameter model trained on thousands of pairs: an AI draft in, the human original out. It rewrites one sentence at a time while a detector checks the whole text, and it refuses any rewrite that adds a number the draft didn't have.
- Real posts, by people
- Flagged as AI0.7%
- AI drafts (Qwen3.5, never trained on)
- Flagged as AI57.3%
- After one rewrite
- Flagged as AI2.3%
- After the best of four
- Flagged as AI0.3%
| 300 held-out LinkedIn posts | Flagged as AI |
|---|---|
| Real posts, by people | 0.7% |
| AI drafts (Qwen3.5, never trained on) | 57.3% |
| After one rewrite | 2.3% |
| After the best of four | 0.3% |
Round 3, report_v3.txt. The flag line is set so 1 in 100 real posts is flagged by mistake. The detector is MELD, an open-source model, and it also chose the best of four, so the one-try row is the fair one. Outside detectors haven't checked these rewrites yet.
Questions
Do AI humanizers actually work?+
Some move a detector's score and some don't. Word-swapping barely does: in our test, cleaning stock phrases took the flagged rate from 93.3% to 92.7%. Across about 20 commercial humanizers in two public benchmarks, the share GPTZero let through ranged from 0% to 88%.
Which AI humanizer is best?+
It depends on which detector you care about, because they disagree. In the public benchmark data, one tool passed MELD about 70% of the time and GPTZero about 11–22%. Test on your own drafts, with a detector the tool doesn't advertise.
Can a humanizer change my meaning?+
Yes. The easiest way to fool a detector is to write something different. Ours keeps meaning similarity above a set bar and blocks new numbers, and it still once added a job title. Read the output against your draft.
Is a humanizer the same as a paraphraser?+
A paraphraser rewords. A humanizer is aimed at a detector score, and the better ones rewrite sentence structure and rhythm. A light paraphrase moved our detector's flagged rate by less than one point.
Is using an AI humanizer allowed?+
For your own posts, emails and messages, yes. Where a school or employer bans AI help, a humanizer doesn't change the rule.
Can I use Patois now?+
Not yet. Patois is in testing. The free AI phrase checker on this site is available now and fixes the surface habits in your browser.
Sources and dates
- MELD, the open-source detector we test with (Hugging Face)huggingface.co
- HumanizerBench (Hugging Face)huggingface.co
- ai-humanizer-benchmark (Hugging Face)huggingface.co
Facts about other products come only from these pages, on the dates shown; prices are as published that day. Macaron's own details are as shipped on October 1, 2026. Products change, so check their sites for the latest.