AI humanizer red flags: the claims vendors make and what the studies allow
Short answer
Seven claims recur on AI humanizer pages read on 9 November 2026, from "We beat every AI detector" to user counts in the millions, and none comes with a dataset or a date. The studies found paraphrased AI text still identified 26 percent of the time by 14 detectors in 2023, so each claim is a red flag.
Seven claims recur across the AI humanizer pages this site read on 9 November 2026. None of the seven arrives with a dataset, a detector version or a date. Searches for AI humanizer scams usually come from someone who has met one of them and wants to know whether the product behind it is real. Mostly it is: the tools rewrite text as described. The red flag is the claim, and this guide quotes each one as the vendor words it, then sets it beside what the published studies allow. The tools themselves are reviewed on the same fields in the reviews hub.
What do AI humanizer pages actually claim?
Each quotation below is the vendor's own wording on the date read, kept under fifteen words so it can be checked against the page.
| Claim type | Vendor and wording, as read 9 November 2026 | What is missing from the page |
|---|---|---|
| Every detector | Grubby AI homepage: "We beat every AI detector" | Which versions, what inputs, what date |
| Every detector | StealthGPT pricing page: "We beat every AI Detector" and "Beat AI detectors and write freely" | Any figure at all beside the logo row |
| All detectors, named | Humanize AI Pro homepage: "We bypass ALL detectors, including Originality.ai, GPTZero, Turnitin, etc." | Turnitin's detector is not sold to the public, so no vendor can test against it directly |
| Human score | Humanize AI Pro homepage: "Aims for a 100% human score" | A score on any named detector, on any date |
| Human score | Undetectable AI homepage: "scoring as human on every major detector" | Which detectors are major, and the readings |
| User count | Undetectable AI homepage: "Trusted by 24M+ users" with a weekly new-user figure | How a user is counted |
| User count | Grubby AI homepage: "Over 1.6 million students have used the Grubby AI humanizer tool" | How a student is counted, or over what period |
| Watermark | Grubby AI homepage: "ChatGPT embeds watermarks in its text" | Evidence that the text pasted carried one |
| Guarantee | Grubby AI homepage: "guaranteed to outperform any competitor" | The test, the competitors, the measure |
| Press | Undetectable AI homepage: "Rated #1 Best AI Detector by Forbes!" | The article, its date, its method |
Two of these vendors sell plans this site has priced: Undetectable AI's entry plan is 10,000 words a month at $5.00 a month charged $60.00 yearly, or $9.99 monthly (Undetectable AI, pricing page, read 9 November 2026), and StealthGPT's Pro plan showed as $1.00 for three days and then $299.99 a year on its pricing page the same day (StealthGPT, pricing, read 9 November 2026). Neither page is a scam in the legal sense. Both make claims that cannot be checked from the page, which is the point.
What do the studies allow a humanizer to claim?
Two 2023 papers set the ceiling, and the ceiling is lower than any headline above.
Weber-Wulff and colleagues tested 14 detection tools on AI text in several conditions. On AI text that had been machine paraphrased, the detectors were still correct 26 percent of the time; on untouched AI text, 74 percent; on AI text a person had edited by swapping synonyms and reordering, 42 percent (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). Paraphrasing moved the readings a long way. It did not move them to zero on any detector in the study.
Sadasivan and colleagues went further and asked whether reliable detection is possible at all. Their paper shows paraphrasing attacks breaking several detectors, including watermark-based ones, and argues from a theoretical bound that as language models improve, even the best detector's advantage over random guessing shrinks (Sadasivan et al. 2023, arXiv, read 9 November 2026). Vendors sometimes cite that paper as proof that humanizers win. The same paper also shows the reverse: that a detector can be built to catch a given paraphraser once it has seen enough of its output, which is what the detector makers have spent the years since doing.
| Vendor claim | What the two studies support | What they do not support |
|---|---|---|
| Passes every detector | Paraphrasing lowers readings on the detectors tested, on the date tested | A pass on every detector, or on any detector at a later date |
| 100 percent human score | Some outputs read as human on some detectors | A guaranteed score, or the same score after the detector updates |
| A clean Turnitin result | Nothing: Turnitin was in the Weber-Wulff set but no public tool can test against it on demand | Any direct claim about a detector the vendor cannot access |
| Removes watermarks | Sadasivan and colleagues weakened a research watermark by paraphrasing | That the pasted text carried a watermark in the first place |
The honest version of the vendor claim is a sentence no marketing page prints: a rewrite changed the readings on these detectors, on this date, at this cost in meaning. The do AI humanizers work guide sets out the full evidence; this page is only about the gap between the evidence and the headline.
Which claims should you treat as red flags before paying?
Ranked by how much money the claim can cost you, not by how loud it is.
| Rank | Red flag | Why it costs money | Where to check instead |
|---|---|---|---|
| 1 | A money-back guarantee tied to a detector the vendor picks | The vendor judges its own guarantee; StealthGPT's terms exclude any result from Originality.ai and accept Turnitin screenshots, read 9 November 2026 | The refund policy page, not the pricing card |
| 2 | A yearly charge shown as a monthly price | Undetectable AI's $5.00 card is a $60.00 charge, read 9 November 2026 | The billing toggle and the "charged" line |
| 3 | "Every detector" or "all detectors" | Nothing supports it; the studies cap paraphrase effects well short of certainty | Two detectors you do not pay the vendor for, on your own text |
| 4 | A user count with no definition | It cannot be checked and says nothing about output quality | Ignore it |
| 5 | Detector logos in a row | Implies a pass without stating one | The absence of a figure beside the logos |
| 6 | "Removes ChatGPT watermarks" | Sells a fix for a mark the vendor has not shown is present | Published statements by the model makers on what they watermark |
| 7 | A press badge | A review on a stated date is not an audit | The article itself, dated |
Rank 1 is the one that turns a marketing claim into a contract term. StealthGPT's terms of service, updated 20 October 2026 and read 9 November 2026, state that all purchases are non-refundable and, separately, that a user whose output is flagged can claim a refund of the current billing cycle on proof from an accepted detector, with results from Originality.ai excluded. The vendor chooses the detector list and judges the proof. Five vendors on this site frame refunds this way, and the terms are compared in AI humanizer refund policies.
Are AI humanizer scams the same thing as bad humanizers?
No, and the distinction matters for what you do next. A scam takes money for nothing or for something other than what was sold. Every tool this site has priced delivers a rewrite; what the pages oversell is the result. A bad humanizer, by this site's method, is one that drops claims, adds errors, or splits your document into pieces that no longer agree, and those faults appear on tools with modest pages and tools with loud ones alike.
The consumer rule that bears on the first kind is the FTC's "click-to-cancel" rule, announced on 16 October 2024, which would require clear disclosure of material terms before billing, express informed consent to recurring charges, and a cancellation route as simple as the sign-up (FTC press release, 16 October 2024, read 9 November 2026). The rule has been revised since to conform to court decisions and is the subject of a 2026 notice seeking comment on further amendments, so its text is in motion; its stated standard, that cancelling should be no harder than signing up, is the one to hold a vendor to. A humanizer that buries cancellation in a support inbox is not thereby a scam, but it has failed that standard on its face.
No detector score should be the only evidence in a misconduct case, and no vendor's detector claim should be the only evidence for a purchase.
What should you do when you meet one of these claims?
Read past it to three pages: pricing with the toggle set to monthly, the refund or terms page, and the privacy policy. Then run the free allowance on a passage of your own and read the result against the input. The 10-question checklist on this site puts those steps in order, and the guides hub carries the read date on every page so a quotation here can be checked against the vendor's page on the day you visit it. A vendor page can change overnight. The studies it has to answer to do not.
Sources
Questions about this guide
Are AI humanizers a scam?
Most are working software with a marketing problem: the tools rewrite text as described, but the pages promise results on every detector and publish no test, no dataset and no date, so the claim is the red flag, not the rewrite.
Can any humanizer make text 100 percent undetectable?
No: Weber-Wulff and colleagues found in 2023 that 14 detectors still identified machine-paraphrased AI text in 26 percent of cases, Sadasivan and colleagues argued the same year that reliable detection and reliable evasion both have limits, and detectors have updated since.
What does a humanizer money-back guarantee usually mean?
A bet on a detector the vendor chooses, judged by the vendor: StealthGPT's terms state all purchases are non-refundable while a separate policy refunds a billing cycle on screenshot proof from an accepted detector, read 9 November 2026.
Is a 'trusted by 24 million users' claim verifiable?
Not from the page: Undetectable AI's homepage states the figure with a weekly new-user count beside it, read 9 November 2026, and defines neither, so the number cannot be checked and should not decide a purchase.
Why do humanizer sites show detector logos?
To imply a pass on each, without stating one: Grubby AI's and StealthGPT's pages show rows of detector logos under the words "We beat every AI detector", read 9 November 2026, and neither page gives a dataset, a date or a detector version.
What should I check before paying a humanizer that makes these claims?
The refund page rather than the headline, the privacy page for training terms, the request cap on the plan you would buy, and a free run of your own 400-word passage read back against the input; the 10-question checklist on this site walks through each.
Related reading
AI humanizer trends: watermarks arrive, paraphrase stays the open problemAI humanizer trends from primary sources: two model makers document text watermarks, detectors sell.
AI humanizer reviewsAI humanizer reviews on the same fields: verdict, best for, main limitation, plan limits and price checked on.
Do AI humanizers work?Do AI humanizers work? What four published studies found on paraphrased text, what GPTZero, Originality.ai.
AI humanizer refund policies: what each vendor's terms stateAI humanizer refund terms from 12 vendors' own pages: who states no refunds, who gives a 3 to 30 day window.
How to choose an AI humanizer: a 10-question checklistHow to choose an AI humanizer in 10 questions: request cap, monthly pool, free allowance, modes, privacy.