Is there an AI humanizer that actually works?

A humanizer has three jobs, and most claims only mention one. The site's protocol scores all three, with weights published before any run.

A tool that lowers a detector reading by swapping words until the sentence no longer says what you meant has done the third job by failing the first. That is why the question "does it work" cannot be answered from a detector screenshot, which is the only evidence most Reddit lists and vendor pages offer. The do AI humanizers work guide covers the research in depth; this page is about choosing.

The one thing every vendor publishes is the set of controls: modes, request caps and plan sizes. These are the levers a user has, so they are worth comparing even before a run.

Two vendors stand apart in what they claim. Grammarly's product page states its humanizer "is not intended to bypass AI detectors" and works on clarity instead (Grammarly AI Humanizer, read 9 November 2026). QuillBot's page describes its naturalness check as one input among several, not a verdict, read 9 November 2026. Every other tool sells the detector outcome and publishes nothing about the meaning cost.

The strength ladders matter because they are the only control that trades the three jobs against each other. Walter Writes AI publishes three modes and a 750-word request cap on its Starter plan, checked on 8 November 2026, so a user can pick how far from the input the output may move; the cap means a 1,500-word essay is two or three requests, rewritten without sight of each other. HIX Bypass, Phrasly and BypassGPT publish three modes each on the same date. A ladder is a promise of control, not of quality; the run tests whether the strongest mode costs more meaning than the default.

The research is about the category, not about products, and it measured only the third job.

Krishna and colleagues reported that their paraphraser, DIPPER, cut DetectGPT's detection rate from 70.3 percent to 4.6 percent at a 1 percent false positive rate (Krishna et al. 2023, arXiv 2303.13408, read 9 November 2026). Weber-Wulff and colleagues found all 14 detectors below 80 percent accuracy, with obfuscation worsening them (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). Sadasivan and colleagues argued that paraphrasing attacks lower detector accuracy and set out why detection has theoretical limits (Sadasivan et al. 2023, arXiv 2303.11156, read 9 November 2026).

The detector side has moved since 2023. Turnitin states that its indicator covers "AI paraphraser or bypasser (also called humanizers) tools", that it trains against "leading paraphraser and bypasser tools" without naming them, and that it holds false positives under 1 percent for documents over 20 percent AI (Turnitin, AI writing detection FAQs, read 9 November 2026). So a 2023 result about a 2023 detector is a finding about the category's past, not a prediction about any tool today.

Here is what can be said from vendor pages and studies alone, by the situation a reader is in. The run's meaning-kept and detector columns will be added to this table when the first run publishes, with the run ID beside them.

Choose by the job you cannot afford to fail. A student whose argument must survive should weight meaning first and distrust any tool whose only evidence is a detector screenshot. A marketer who can proofread 500 words can tolerate more errors and care more about readability. Anyone in a graded course should read their institution's AI policy first; a humanizer does not make undisclosed AI use acceptable, and this site does not help anyone misrepresent authorship.

The protocol this site uses scales down to a single check.

1. Take one text of about 400 words that you wrote with AI help and know well.
2. Run it through the tool's default mode once. No retries, no cherry-picking.
3. Read output against input and mark every claim that is missing, changed or invented. One invented fact is a fail for most purposes.
4. Count grammar, wrong-word and broken-sentence errors per 100 words.
5. Check the input and the output on two public detectors the same day, and write down the date; detector models change without notice.
6. If the tool offers a stronger mode, repeat and compare the meaning cost, not just the reading.

No detector score should be the only evidence in a misconduct case. Dated runs under the full protocol, 24 inputs, 6 human controls, five detectors and two blind meaning ratings, will be listed in the runs archive with every raw file, and the full list carries each tool on the same fields. The rest of the series is in the guides hub.