How to choose an AI humanizer: a 10-question checklist

Work down the table in order. The first seven can be settled in ten minutes on the vendor's pages; the last three need the free allowance and a passage of your own.

Because the cap decides how many pieces your document becomes, and the pieces are where humanizers do their quiet damage. A 2,000-word report on a 750-word cap is three requests; each is rewritten without sight of the others, so a term defined in piece one can come back as a synonym in piece three. The price per word is the same either way, which is why price sits at question 3 and the cap at question 1.

Caps are the vendors' own figures from their pricing pages. The complete list of caps, free and paid, is on the humanizer word limits guide, and the arithmetic on what each plan costs per 10,000 words is in AI humanizer pricing compared.

Three answers end the conversation whatever the price. A privacy page that says your text trains the model with no opt-out, if the text is confidential: WriteHuman's privacy notice states "There is no setting to opt out of this use", read 9 November 2026. A refund term that is a detector bet rather than a window, if you want the right to change your mind: StealthGPT's terms of service state that "All purchases are non-refundable" alongside a separate detector-tied policy (StealthGPT, terms of service, read 9 November 2026). And a homepage that promises a result on every detector, because no study supports it.

On that last point the evidence is specific. Weber-Wulff and colleagues ran 14 detectors over AI text that had been machine paraphrased and found the detectors still identified it correctly 26 percent of the time, the lowest rate in their study but not zero (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). Turnitin, for its part, lists "text that has been paraphrased without developing new ideas" among the patterns that can produce a false positive (Turnitin, AI writing detection FAQ, read 9 November 2026). A vendor claiming a clean result on every detector is claiming something the detector makers and the researchers both say is not how detection works. The patterns to watch for are set out in AI humanizer red flags.

Use the free allowance on a passage you wrote or know well, 300 to 500 words, with at least one number, one qualifier and one technical term in it. Run it once on the default mode. Then do three things.

That is a smaller version of what the published method does with 24 inputs, 6 human controls and 5 named detectors, and it is enough to rule a tool out. It is not enough to rule one in, because detector models change without notice and a reading is true for its date. No detector score should be the only evidence in a misconduct case, and no single pass should be the only reason to pay for a year.

Question 7 has a legal floor in the United States. The FTC announced a final "click-to-cancel" rule on 16 October 2024 requiring sellers to disclose material terms before taking billing details, obtain express informed consent to recurring charges, and provide a cancellation mechanism as simple as the sign-up (FTC press release, 16 October 2024, read 9 November 2026). The rule has since been revised to conform to court decisions and is the subject of a 2026 notice seeking comment on new amendments, so the 2024 text is not the text in force; what it tells a buyer is which cancellation routes meet the agency's stated standard and which run through a support inbox. The vendor-by-vendor terms are in AI humanizer refund policies.

Outside the United States, statutory rights differ by country and are not covered here. Wherever you are, the practical rule is the same: put the renewal date in a calendar the day you pay, because every vendor this site has read says cancellation takes effect at the end of the paid term.

The checklist gives you a shortlist, not a winner. If your documents are under 750 words and you want a strength setting, the shortlist is one shape; if you paste 3,000-word chapters, it is another; if your text is confidential, two tools fall off it before price is discussed. The ranking on bestaihumanizer.ai applies the same ten questions to every tool on the same fields, and each review carries a critical main limitation rather than a recommendation. Price is reported there and never scored. The guides hub holds the pages that go deeper on each question, each with the date its figures were read.