Do AI Humanizers Work? We Tested Them Against 4 Detectors (2026 Data)
Every humanizer on the market, including ours, promises to make AI text pass detection. Customers kept telling us it was not happening, so in September 2026 we measured it: eight AI-written texts, two humanizer modes, and four detectors. Here are the numbers, including the ones that embarrassed us.
Short answer
Humanizers work against pattern-based detectors and mostly fail against classifier-based ones. In our test a humanized news article dropped from 72 to 12 on our detector and from 100 to 63 on ZeroGPT, while GPTZero and Pangram scored every version 100% AI. A light one-click rewrite can even push a score up.
How we tested
- Texts: eight AI-written samples from seven models (GPT-4o, GPT-4o-mini, GPT-5.1, Gemini 2.5, Claude 4.5, DeepSeek, Llama 3.3) across a news article, a student essay, a cover letter, product copy, a tech explainer, a blog post, an advice answer and a personal reflection. Two human-written controls scored 24 and 32.
- Humanizer: our AI humanizer in its default mode and in the aggressive Pro-only mode, which also inserts a few deliberate typos.
- Detectors: our own AI detector via the API, then the same texts pasted by hand into the GPTZero, ZeroGPT and Pangram web apps. Pangram was only run on the news article. All numbers are the “% AI” each tool displayed.
Result 1: what a humanizer did to our own detector’s score
| Sample | Original | Default humanize | Aggressive mode |
|---|---|---|---|
| GPT-4o news article | 81 | 36 | 25 |
| GPT-4o-mini student essay | 91 | 77 | 25 |
| Gemini 2.5 cover letter | 65 | 49 | 35 |
| Llama 3.3 product copy | 49 | 47 | 35 |
| DeepSeek tech explainer | 30 | 37 | 15 |
| GPT-5.1 blog post | 20 | 25 | 22 |
| GPT-5.1 advice answer | 20 | 43 | 25 |
| Claude 4.5 reflection | 15 | 59 | 15 |
Three things stand out. The aggressive mode lowered every score and got all eight texts under the 50 line. The default mode, the one most people use, pushed three of eight texts up: the Claude reflection went from 15 to 59 and the GPT-5.1 advice answer from 20 to 43. And half of the raw AI texts already scored as human before any humanizing, which says as much about detector limits as it does about humanizers.
The reason for the “went up” cases was easy to find once we looked at the diffs. The light rewrite kept re-inserting exactly the vocabulary detectors are tuned for: “amidst”, “bustling”, “subsequently”, “delve”. A rewrite that adds AI tells is worse than no rewrite.
Result 2: the same texts on GPTZero, ZeroGPT and Pangram
| Text | Ours | GPTZero | ZeroGPT | Pangram |
|---|---|---|---|---|
| News article, original | 72 | 100 | 100 | 100 |
| News article, default humanize | 51 | 100 | 100 | 100 |
| News article, aggressive mode | 12 | 100 | 63 | 100 |
| Student essay, original | 91 | 100 | 100 | not run |
| Student essay, default humanize | 77 | 100 | 94 | not run |
| Student essay, aggressive mode | 25 | 100 | 54 | not run |
This is the table that answers the question. ZeroGPT, which leans on statistical patterns, moved with the rewrite: 100 to 63 on the article, 100 to 54 on the essay. GPTZero and Pangram, which use trained classifiers and explicitly train on humanizer output, did not move at all. Every version came back 100% with high confidence.
So “do AI humanizers work?” has two honest answers. Against a detector that counts patterns, yes, often dramatically. Against a detector trained to recognise humanized text, no, not with a one-click rewrite, and we could not find a mode that changed that.
Result 3: what real users see
We also looked at 30 days of anonymised scan history on our own site. Users re-scanned a related piece of text within 20 minutes 1,125 times, which is the “humanize, then check again” loop in the wild. The median change in score was zero. Of the 304 re-scans that started at 50 or higher, only 105 finished below 50. That matched the customer complaints, and it is why we changed the tool.
What we changed after this test
Two things. Signed-in rewrites now use a larger rewrite model, and every rewrite is checked by the detector before it is returned; if the output still scores above the target, the tool rewrites once more with the flagged sentences as feedback. On the same style of corpus (24 AI texts, scored by our detector) the numbers moved from this:
| Variant | Mean AI score | Flagged (50+) | Score went up |
|---|---|---|---|
| Original AI text | 45.1 | 7 / 24 | – |
| Old default rewrite | 42.0 | 6 / 24 | 11 / 24 |
| New default (model + detector check) | 27.8 | 1 / 24 | 3 / 24 |
| Aggressive mode | 30.8 | 2 / 24 | 6 / 24 |
That is measured on our own detector, which is the fair caveat. We are not claiming those numbers transfer to GPTZero or Pangram, because the table above says they do not. The humanizer now also shows you the before-and-after detector score inline, so you know what you are holding before you paste it anywhere.
Why the aggressive mode is not a free win
The mode that beat ZeroGPT does it partly by roughing up the text: it inserted typos such as “netx” and “wll”, added invented hedges like “or maybe it was the week after, I forget”, and cut the student essay from 254 words to 160. For a forum post that may be fine. For an essay, a cover letter or anything with your name on it, you would have to fix all of that by hand, at which point you are doing the real revision anyway.
When a humanizer is worth using
- Polishing your own draft. You wrote it, AI tightened it, and it now reads stiff. A humanizer plus a quick edit is the fastest fix, and there is nothing to hide.
- Marketing and web copy that will be judged by readers and by pattern-based filters, not by a classifier with a stake in the outcome.
- Not for graded work where the writing has to be yours. Detectors like Turnitin train on humanized text, institutions increasingly ask for drafts, and a rewrite tool does not change who did the thinking. If disclosure is required, disclose.
How to check whether a humanizer actually worked
- Run the original through a detector and note which sentences are highlighted.
- Humanize it, then run the output through the same detector and a second one.
- Read the highlighted sentences, not the headline number. Those are the lines to rewrite yourself.
- Add one specific detail per paragraph that only you could know. That moves classifiers more than any rewrite.
- Re-check. If two detectors disagree, the text is in the mixed zone and needs more of you in it.
The step-by-step version of that edit is in how to humanize AI text. For the tools themselves, see our best AI humanizers of 2026 roundup and the honest review of whether Undetectable AI is legit.
Test your own text
Paste any draft into the free AI detector to see which sentences are flagged, then send it to the humanizer and watch the before-and-after score. No sign-up for the first scan.
Frequently asked questions
Do AI humanizers actually work?
Sometimes, and it depends on the detector. In our September 2026 test a humanizer lowered scores on pattern-based detectors (our own and ZeroGPT) but not on classifier-based ones: GPTZero and Pangram flagged every humanized version at 100%. A one-click rewrite is not a guarantee.
Can AI detectors detect humanized text?
Often, yes. Detectors trained on humanizer output learn the tells humanizers leave behind: odd synonym swaps, flattened rhythm, filler hedges and clipped sentences. That is why the same text can read as human on one detector and 100% AI on another.
Does GPTZero detect humanized text?
In our test it did. Every humanized output we pasted into GPTZero, in both a light mode and an aggressive mode, came back 100% AI with high confidence. Your results may differ by text, but do not assume a humanizer beats GPTZero.
Does Undetectable AI work?
We did not test Undetectable AI in this run, so we cannot give you a number. The mechanism is the same as any rewrite tool, which means the same limits apply: it can move pattern-based scores and is much less likely to move a classifier trained on humanized text. See our separate review of whether Undetectable AI is legit.
Does Turnitin detect humanized text?
Turnitin does not let students run their own checks, so nobody outside an institution can test this directly. Turnitin has said its model is trained to spot paraphrased and rewritten AI text, so treat humanized text as detectable and rewrite in your own words if the work has to be yours.
What actually lowers an AI detection score?
Real revision: restructure the argument, add specific first-hand detail, vary sentence length, and cut the stock phrases. Then re-check with two detectors and read which sentences are still flagged. A humanizer is a starting point for that edit, not a replacement for it.
Method notes: scores are the “% AI” shown by each tool on 12 September 2026. Our own scores come from the production detector API on a Pro account. GPTZero, ZeroGPT and Pangram scores come from pasting the identical text into each site’s free checker. Detectors change their models often, so treat this as a snapshot and re-test before relying on it.