Are AI humanizers safe and legit?

Is AI Humanizer legit?

"AI Humanizer" is a category, not one company. Several products use the name, and dozens more sell the same function. A product is legit as a business if it has a named operator, published terms and privacy policy, working support and clear pricing. Whether it is safe for you is a separate question about data, money and accuracy.

Two signs that call for caution: marketing that promises a guaranteed detector result, and terms that tie your refund to one. No detector, and therefore no humanizer, can guarantee an outcome. Detection vendors retrain against rewriting tools; Turnitin added detection aimed at humanizer-modified text in August 2025. My explainer on how AI detectors work covers why scores move.

Also check that the use itself is allowed. Schools and publishers set their own AI rules, and my guide on whether using an AI humanizer is cheating quotes several real policies.

My take: I judge legitimacy in two minutes. If I can't find the company name, the terms and a support address from the footer, I close the tab. A product that hides who runs it has no reason to look after your text.

Are AI humanizers safe?

Privacy and training on your text

Every humanizer receives the full text you paste. What happens next is set by its privacy policy and terms, and the defaults differ.

• Training defaults. Grammarly's Product Improvement and Training article says the control is on by default for Free, Premium and single-user Pro accounts, meaning content can be used to train models, and that users can switch it off. It is off by default for Education accounts and accounts bought through Grammarly's sales team. That is a clear, documented setting. Many smaller humanizers publish no setting this specific, so I treat silence on training as a yes and keep sensitive text out.
• No opt-out at all. Some vendors say plainly that they train on what you submit. WriteHuman's terms say it trains its models on submitted text with no setting to turn that off, as my WriteHuman AI humanizer review explains. For a tool like that, only public copy belongs in the box.
• License clauses. Undetectable AI's terms define "Contributions" to include text and writings you submit through the services, and grant the company an "unrestricted, unlimited, irrevocable, perpetual" worldwide license to use, copy, publish and distribute them. The same terms say you keep ownership. Read clauses like this as permission the vendor has reserved, whatever its current practice.
• Publisher rules. If you write for journals, Elsevier's generative AI policy tells authors to check a tool's terms so that they grant it only the right to provide the service, not the right to train on their material, and to take particular care with personal data.

My AI tool privacy checklist walks through the same checks for chatbots.

My take: I sort every paste into one of three buckets before it goes anywhere. Public copy (a blog draft, a product description) is fine in almost any tool. Internal copy (a client email, a plan) goes only into a tool with a training toggle switched off. Confidential copy (contracts, medical or HR details, unpublished research) goes into no third-party humanizer at all.

Subscription, cancellation and refund traps

The money risks are in the terms, not the pricing page.

• How you cancel. Undetectable AI's terms, last updated 1 January 2025, say California and New York residents can cancel in the dashboard, while everyone else must email support from the account address. Requests through contact forms, social media, phone calls or chargebacks "do not constitute valid cancellation," and a chargeback does not cancel the subscription. Its help center adds that remaining premium credits expire on cancellation.
• What you get back. The same terms say refunds will not be provided unless required by law. A refund of credits is available only if a submission "provably flags as less than 75% human on a reputable AI detector," requested within 30 days. The terms also say a yearly subscription is valid only while the company operates, with no refund if it stops.
• Non-refundable by default. StealthGPT's terms, last updated 20 October 2026, say "All purchases are non-refundable." You can cancel by logging in, and cancellation takes effect at the end of the paid term. The terms list weekly as well as monthly and yearly plans, so a forgotten weekly plan renews four or five times a month. API purchases are also non-refundable.
• Refund pages that disagree. Humbot's pricing table mentions 30 days, while its Refund Policy sets a 7-day window with strict usage caps, as my Humbot review shows. Natural Write's FAQ promises a 14-day refund that its own Terms of Service rule out, and its site names no company behind it; my Natural Write review quotes both pages. When two vendor pages disagree, I go by the stricter one.

None of this is unusual for small software subscriptions. It does mean the safest approach is to start on a free tier or the shortest plan, put a cancellation reminder in your calendar, and cancel exactly as the terms describe.

My take: if a tool's terms make cancellation harder than signing up, I pay monthly at most, never yearly, and I screenshot the cancellation email the day I send it.

Meaning drift

The most common harm is quieter than a billing dispute: the rewrite says something different from what you meant.

Masrour and colleagues audited 19 humanizer and paraphrasing tools, including Undetectable AI, StealthGPT, StealthWriter, WriteHuman, Phrasly and QuillBot, in a paper presented at GenAIDetect 2025. They sorted the tools into three tiers. The best rewrote text while keeping tone, vocabulary level and complexity. Middle-tier tools kept the message but degraded quality. The weakest added hallucinated citations, stray characters and nonsensical phrases, and often distorted meaning. In a GPT-4o fluency comparison, even the best tier was preferred over the original AI text only 26% of the time; the weakest, 2.67%. The authors work at Pangram Labs, a detector company.

Practical checks against drift:

• A hedge removed ("may" became "does").
• A number rounded or changed.
• A citation that was not in your draft.
• A shift in register, for example a casual email turned academic.

Example (I wrote this to show the pattern; it did not come from any tool):

• Original: "Early results suggest the new dosage may reduce symptoms in some adult patients (Lee, 2024)."
• Drifted rewrite: "Research proves the new dosage reduces symptoms in patients, according to several studies (Lee, 2024; Patel, 2023)."

Four things changed in one sentence. "Suggest" became "proves", "may" disappeared, "some adult patients" widened to all patients, and a second citation appeared that was never in the draft. Each one reads smoothly, which is exactly why drift is dangerous: nothing looks broken.

My guide to testing rewrite quality has a full side-by-side method.

Is Undetectable AI legit?

Undetectable AI is an operating business with published terms, a help center and a support address, so it is legitimate in that sense. My Undetectable AI review covers its plans and cancellation steps. Its terms are also among the stricter ones in this market on cancellation and refunds, as set out above, and its refund condition is tied to a detector score. On whether it does what it says, the 2025 audit above included it among the 19 tools, and research on detection shows results after rewriting shift in both directions. Treat any vendor's detector claim as the vendor's claim.

My take: I'd call Undetectable AI legit but unforgiving. If you subscribe, do it from a monthly plan, use the email cancellation route the terms require, and don't count on a refund. The same reading applies to StealthGPT, whose terms I go through in my StealthGPT review.

Red flags in humanizer terms

I read a lot of humanizer terms, and the same clauses keep turning up. None of them is illegal, and several are common across software. Together they tell you how much the vendor has protected itself at your expense.

The first two rows are about your text. The middle five are about your money. The last two are about honesty. A tool that trips one row in each group is a tool I would use only for throwaway copy.

How I read a privacy policy in five minutes

You don't need to read every word. Search the page (Ctrl+F or Cmd+F) for these terms and read the paragraph around each hit:

1. "train" or "improve": does submitted content feed models, and can you opt out?
2. "license": is it limited to providing the service, or perpetual and worldwide?
3. "retain" or "delete": how long is text kept, and can you ask for deletion?
4. "third part": which processors receive your text, such as model providers?
5. "sell" or "share": is personal data passed on for advertising?

If a policy has none of these words, it isn't a policy you can rely on. That is the moment I stop and pick a different tool.

Risk by use case

The same tool can be fine for one job and a bad idea for another. This is how I'd weigh the common uses.

My take: the risk rises with two things, how private the text is and how much rides on its accuracy. If both are high, no humanizer is the right tool.

Safer ways to get the same result

If what you want is a draft that reads like you, a paid humanizer is only one route, and often not the safest.

• Edit by hand. Nothing leaves your machine. My guide to how to humanize AI text by hand sets out a seven-step method.
• Prompt a chatbot you already trust. If your company already approves ChatGPT or Claude, you avoid adding a new vendor. My ChatGPT humanizer prompts show how to use a sample of your writing as the instruction.
• Use an open prompt you can read. The Claude humanizer skill guide covers blader/humanizer, an MIT-licensed skill whose full instructions are public.
• Pick a tool with documented controls. Grammarly's training toggle is the kind of setting I look for. My Grammarly AI Humanizer review covers that product, and the best AI humanizers ranking and free AI humanizers list compare others on privacy and refund terms alongside price.

If something goes wrong

Problems with humanizers fall into a few predictable kinds. Here is what I'd do first, and what to do if that fails.

Keeping your drafts is the single habit that helps in almost every row. Version history in Google Docs or Word costs nothing and settles questions fast.

A checklist before you use a humanizer

1. Is the use allowed? Check your school, employer or publisher's AI rules first.
2. Who runs it? Find a company name, address and support contact.
3. What happens to your text? Look for training, retention and license clauses. Skip tools with no privacy policy.
4. Can you switch training off? If there is a toggle, use it.
5. How do you cancel? Note the method the terms require, not the one you assume.
6. What is refundable? Assume nothing is unless the terms say so.
7. What is the renewal interval? Weekly plans add up quickly.
8. Test on throwaway text. Use a paragraph with no personal data.
9. Check the output line by line against the original for facts, hedges and citations.
10. Keep your drafts. Version history protects you if your own writing is ever questioned.

What the research says

Three findings recur across independent papers.

Rewriting lowers many detectors' hit rates, often with some quality cost. Sadasivan and colleagues, in work published in TMLR, found recursive paraphrasing significantly reduced detection rates while only slightly degrading text quality in many cases. Cheng and colleagues reported at NeurIPS 2025 in Adversarial Paraphrasing that simple paraphrasing actually raised detection on two detectors they tested, while detector-guided paraphrasing lowered it, with "mostly a slight degradation in text quality."

Detectors can be retrained against humanizers. The 2025 audit's own detector, trained with a small amount of humanized text, caught 98.26% of humanized AI essays at a 5% false-positive rate in the authors' benchmark. That is a vendor-authored result on its own test set.

Detectors are not reliable enough to settle cases. Perkins and colleagues found in a 2024 study of six detectors that accuracy fell from 39.5% to 17.4% on manipulated content, and concluded the tools cannot currently be recommended for deciding integrity violations. That is also why I tell readers not to buy a humanizer on the strength of a detector promise: the scores on both sides are too unstable to bet a grade or a job on. My guide on whether Turnitin can detect humanized AI text covers Turnitin specifically.

For the full research picture, read my guide to do AI humanizers work, and for what schools check beyond a score, my guide on whether teachers can detect AI.

How we checked this

• Read on 29 October 2026: Undetectable AI's terms (last updated 1 January 2025) and its cancellation help article; StealthGPT's terms (last updated 20 October 2026); Grammarly's Product Improvement and Training article; Elsevier's generative AI policies for journals; Masrour et al. 2025 on ACL Anthology; Sadasivan et al. (arXiv 2303.11156, TMLR); Cheng et al. 2025 (arXiv 2506.07001, NeurIPS 2025); Perkins et al. 2024 (arXiv 2403.19148).
• What I did: I read each vendor's terms and help pages, compared the cancellation and refund clauses side by side, and read the papers in full.
• Scope: documented terms and named research. Terms change; the "last updated" dates above show which version was read.

More evaluation guides are in the HumanRank guides hub.

Frequently asked questions

Are AI humanizers safe to use? They can be, for low-stakes text, if you have read the privacy, cancellation and refund terms and you check the output. The main risks are training or broad licenses on your text, strict refund rules and rewrites that change meaning.

Is AI Humanizer legit? "AI Humanizer" covers many products. Look for a named company, published terms and privacy policy, working support and no guaranteed detector result.

Is Undetectable AI legit? Yes, it is a real company with public terms and support. Read those terms first: outside California and New York you cancel by email, and refunds come only where the law requires.

Do AI humanizers store or train on my text? It depends on the vendor. Grammarly documents a training toggle that is on by default for individual accounts. Some humanizer terms reserve broad, perpetual licenses over submitted content. Check the policy for each tool.

Can I get a refund from an AI humanizer? Usually not. StealthGPT sells nothing refundable, and Undetectable AI refunds only when the law forces it, plus a small credit refund linked to a detector score.

Can an AI humanizer change the meaning of my text? Yes. A 2025 audit of 19 tools found the weakest added fake citations and nonsensical phrases and distorted meaning. Compare the output with your original every time.

Is it safe to use a humanizer for school work? Only if your course allows AI help, and usually with disclosure. Schools set their own rules, and a rewrite does not change what those rules cover.

REVIEW NOTEI checked this on 29 October 2026 against the published terms of two popular humanizers, Grammarly's training controls, a journal publisher's AI policy and four research papers on humanizing and detection.