How I test AI humanizers
Ten tests, the same for every humanizer, each graded from 1 to 5 on a fixed scale. The overall grade is their average plus my verdict, which can move it by up to one point. Every test is run by hand on an account I use, and the original passages, the rewrites, the detector verdicts, the counts, screenshots and timings behind each grade go in the full test results published with the review.
Tested and written by Miles Foster, Owner and tester, RankDude.

Each humanizer gets the same corpus, built before any tool is opened: 50 AI-written passages of about 300 words each, five from each of ten prompts and spread across six models (GPT-5, GPT-3.5, Claude, Gemini, Llama and DeepSeek) at default settings, every one checked first so that all four detectors call it AI before it is touched, leaving the humanizer something real to beat. Thirty of those passages carry a set of checkable facts written into them in advance (names, dates, figures, a stated claim and its reason) so a rewrite can be marked for whether it kept the meaning or garbled it. A further 24 passages, twelve to be humanized and twelve human controls, run across Spanish, French, German and Chinese for the languages test. The verdict on every rewrite is taken from the same four fixed detectors used in the AI-detector reviews (GPTZero, Originality.ai, Copyleaks and Turnitin) and one fixed plagiarism checker, all named with their version on the test date, so a rewrite that beats one detector but not another is counted as exactly that rather than waved through.
Each humanizer is run on the best plan the account offers, on a free or trial account where one exists, at the default settings and the strongest humanizing mode it ships with, with each passage pasted or uploaded once and the first rewrite taken: no retries, no re-rolling, no picking the best of several. Every rewrite is then submitted once to each of the four detectors and the plagiarism checker and the first verdict is taken, with a detector counted as beaten only when it reads the rewrite as human (50% or below on a tool that returns a probability, the threshold recorded). Nothing is graded from a vendor's bypass claim, a benchmark table or a marketing page. Each review states how the account was obtained: used on a free tier, bought at the public price, or provided by the vendor. A vendor-provided account never changes a grade, and a vendor never sees a review before it is published.
How I rate
Every tool gets one score out of 5.
First, I run the same 10 tests on every tool and give each one a grade from 1 to 5. The average of the graded tests is the starting point; a test that cannot be graded is left out of it, and the review says how many tests the score is based on.
Then I add my verdict: what the tests cannot measure, such as how the tool holds up after the test weeks, how billing and cancelling work in practice, and how it compares with the others in the category. My verdict can move the score up or down by one point at most.
Every review shows how much my verdict changed the score, and why, in its scores table.
What a grade means
| 5 | Excellent |
| 4 | Good |
| 3 | Average |
| 2 | Below average |
| 1 | Poor |
The final score is the average of the graded tests (ten when every test can be graded), to one decimal place, plus my verdict, which can move it by up to one point. If a test couldn't be finished properly it gets no grade, and the review says how many tests the score is based on. A humanizer with no file upload, no bulk mode or no non-English support is not penalised on a capability it does not claim; that test is scored on what it can attempt, with the count stated, or left out and noted. A humanizer that beats every detector but leaves the meaning wrecked or the text unreadable is marked down hard, because a rewrite that no one can use is worth nothing however well it bypasses. Some tests on a free plan use fewer passages than usual because of daily or word limits; I run them again on a paid plan when I can.
The 10 tests
1.Detection-Bypass Rate
This is the headline test, because beating detectors is the whole promise of the product. Each of the 50 AI passages, every one confirmed AI by all four detectors before humanizing, is rewritten once on the strongest mode the plan offers and the first rewrite is kept. That rewrite is submitted once to GPTZero, Originality.ai, Copyleaks and Turnitin, the same detectors and thresholds the AI-detector reviews use, and the first verdict is taken: a passage counts as bypassed only when all four call it human (a probability of 50% or below where a detector returns one, the threshold recorded). The full bypass count out of 50 is the grade, and the share that beats some but not all detectors is reported, because a rewrite that slips past one checker and is caught by the next has not done its job.
- 5: 44 to 50 of 50 read as human by all four detectors
- 4: 37 to 43 bypassed
- 3: 28 to 36 bypassed
- 2: 15 to 27 bypassed
- 1: fewer than 15 bypassed
2.Coverage Across Detectors
The results from the bypass test are broken out by detector, so GPTZero, Originality.ai, Copyleaks and Turnitin each get their own bypass rate out of 50. The grade is set by the detector the humanizer is weakest against, not the average, because a user who is checked by Turnitin gains nothing from a tool that only fools ZeroGPT. The per-detector table is published, the detector it beats most and least often is named, and Turnitin and Originality.ai are called out separately since they are the ones academic and publishing readers are checked by. A detector the humanizer's own site says it does not target is still tested and still counted, since the user is checked by it either way.
- 5: every detector beaten on 85% or more
- 4: every detector on 70% or more
- 3: every detector on 55% or more, or one weak with the rest strong
- 2: one detector beaten under half the time
- 1: two or more detectors beaten under half the time
3.Meaning & Fact Preservation
The 30 passages each carry a fixed set of checkable elements written in before any tool is opened: named people, dates, figures, a stated claim and the reason given for it. Each is rewritten once and read against the original line by line. A rewrite counts as clean only when every fact is still correct and the argument still says the same thing; a changed number, a dropped name, a reason turned around, a claim softened into its opposite, or an invented detail is a failure, and the specific breakages are listed. This is the test that separates a usable rewrite from a dangerous one, because a humanizer that sails past every detector while quietly changing a date or a figure hands the user text that is now both undetectable and wrong.
- 5: 28 to 30 of 30 keep every fact and the meaning
- 4: 25 to 27 clean
- 3: 20 to 24 clean
- 2: 14 to 19 clean
- 1: under 14, or facts routinely changed or invented
4.Readability & Grammar
All 50 rewrites are read in full and marked clean or broken. A passage is broken when it carries a clear grammar or spelling error the original did not, a mangled or invented word, a sentence that no longer parses, or a humanizer artefact (a hidden or non-standard character, a zero-width space, a swapped look-alike letter, random capitalisation) planted to fool detectors at the cost of the text. The count of broken passages out of 50 is the grade, and a readability score of the clean rewrites against the originals is reported alongside. A humanizer that defeats detectors by stuffing the text with invisible characters or nonsense words is marked down here, because the user still has to hand that text to a reader.
- 5: at most 2 of 50 broken
- 4: 3 to 6 broken
- 3: 7 to 12 broken
- 2: 13 to 22 broken
- 1: more than 22 broken, or artefacts throughout
5.Natural, Human Voice
Twenty rewrites are shuffled in with ten genuinely human-written passages of the same length and genre and read blind, each marked on whether it reads as something a person would write and whether it still carries the tone, register and purpose the original set out with (an argument still argues, an email still sounds like an email). A rewrite fails when it reads as obviously machine-shuffled, when the voice has flattened into word-salad synonyms, or when the point of the piece has drifted. The share that passes as naturally human and on-intent out of 20 is the grade, and the tells that gave the failures away are described. Beating a detector by making the text read worse than the AI original is treated as a loss, not a win.
- 5: 17 to 20 of 20 read as naturally human and on-intent
- 4: 14 to 16
- 3: 10 to 13
- 2: 6 to 9
- 1: 5 or fewer, or the voice collapses into synonym-salad
6.Plagiarism Introduced
The 50 AI originals are run through the fixed plagiarism checker first and recorded as clean (AI drafts rarely match a source); each rewrite is then run through the same checker once and its match score taken. A rewrite counts against the tool when it newly trips the checker above 10% matched text, because some humanizers reach for stock human phrasing lifted wholesale from the web and so swap an AI flag for a plagiarism flag. The count flagged out of 50, the highest match score seen, and whether the matched text was a real lift or a common phrase are reported. A humanizer that makes text undetectable as AI only by making it detectable as plagiarism has moved the problem, not solved it.
- 5: at most 1 of 50 flagged above 10%
- 4: 2 to 4 flagged
- 3: 5 to 8 flagged
- 2: 9 to 15 flagged
- 1: more than 15 flagged, or lifted text throughout
7.Non-English Support
Three AI passages in each of Spanish, French, German and Chinese are humanized once and run through the detectors that support the language, and each is marked bypassed only when it reads as human and still carries its facts and meaning, judged against the twelve human controls so a rewrite that turns into broken grammar in that language is caught. Both failure kinds are reported: passages still caught as AI, and passages whose meaning or grammar broke in the language. A humanizer that states it is English-only is tested in English alone, marked as not offering non-English, and this test is left out of its score with the fact stated rather than counted against it.
- 5: 10 to 12 of 12 bypass and keep meaning
- 4: 8 or 9
- 3: 6 or 7
- 2: 3 to 5, or grammar breaks across languages
- 1: 2 or fewer, or English-only when it claims otherwise
8.Input, Length & Output Support
Each point is checked by doing it on the test date: paste a passage; upload a PDF or DOCX; humanize several passages in one batch; a per-run word or character cap generous enough for a full essay (2,000 words or more) on the plan tested; a choice of strength or mode (light, strong, an academic or readability setting); keeping the formatting (paragraphs, headings, lists) through the rewrite; a built-in AI detector so the user can check the result in place; a browser extension or word-processor add-in; a documented API; and an export or copy of the rewrite that keeps its formatting. A point counts only when it works end to end on the account used, not when it is merely listed on a plan I am not on, and each is screenshotted.
- 5: all ten points
- 4: 8 or 9
- 3: 6 or 7
- 2: 4 or 5
- 1: 3 or fewer, or paste-only with a tight cap
9.Humanizing Speed
All 50 rewrites in the bypass test are timed on the same connection, from pressing humanize until the finished output is shown; progress bars and queue messages do not count. The median is the grade; the average, fastest, slowest, any failed or timed-out rewrites, and the free-tier queue or slow-down when it differs from the paid tier are reported. Batch and file rewrites are timed separately and noted, since a long document naturally takes longer than a pasted paragraph.
- 5: median under 8 seconds
- 4: 8 to 14.9 seconds
- 3: 15 to 29.9 seconds
- 2: 30 to 59.9 seconds
- 1: 60 seconds or longer, or rewrites time out
10.Pricing & Free-Tier Limits
On the free tier I humanize text until the tool stops me and record the real daily or per-run word limit, whether an account is needed, whether the free tier runs the same engine and strength as the paid one, and whether bulk, file upload, the strongest mode and the built-in detector are free or paid. Every plan is then read from the checkout, not the marketing page, with the date: the monthly cost billed yearly, what the paid plan lifts or adds, any credit or per-word metering, and the cost of humanizing 1,000 words worked out in dollars on the cheapest plan that removes the free cap. Any pay-as-you-go or credit pricing is worked out as the cost of a typical day's rewriting.
- 5: free tier of 500 words a day or more, or a paid plan under $12 a month that lifts limits and adds the strongest mode and bulk
- 4: a usable free tier and a paid plan $12 to $19.99
- 3: a tight free tier, or a paid plan $20 to $29.99
- 2: $30 or more, or no usable free tier
- 1: poor value or unclear credit pricing
Documented but not graded
A built-in AI detector, a bundled plagiarism checker, an AI writer or essay generator, a paraphraser or summariser, a choice of writing modes and tones, invisible-character and watermark removal, a Chrome extension or word-processor add-in, bulk and API access, team and classroom dashboards, a mobile app and languages beyond English are checked and described in every review. When they work well they count in the written assessment, but they never add points to the overall grade.
Rules that apply to every test
- Real use of the product on a real account, never vendor bypass claims, benchmark tables or promotional descriptions.
- Identical passages, detectors, plagiarism checker, thresholds and sample sizes for every humanizer in the category, with the first rewrite and the first detector verdict counted every time and the strongest mode the plan offers.
- The verdict on every rewrite comes from the same four detectors (GPTZero, Originality.ai, Copyleaks, Turnitin) the AI-detector reviews use, run at their defaults with the version recorded, so the two categories stay consistent and a rewrite is judged against the inverse goal.
- Only AI passages I generated myself are humanized, never a real person's work, and the tests measure the tool, not to pass off deceptive text as anyone's own.
- Every grade has evidence: the original, the rewrite, the detector verdicts, a count, a measurement, a screenshot or a table, in the full test results.
- A grade changes only when a retest produces new evidence; the review records the date, the humanizer's version and the detector versions, because both the tools and the detectors change often.
Questions about a specific grade go to the contact page. Makers can submit an AI humanizer for testing. The tested ranking is on the AI humanizers page.