What is an AI humanizer?

You paste text, pick a mode, and a language model rewrites it. The rewrite is aimed at the properties AI detectors score, and the strongest modes rewrite most.

The research behind this is from 2023. Krishna and colleagues built DIPPER, an 11 billion parameter paraphraser that controls lexical diversity and content reordering, and reported that it dropped DetectGPT's detection accuracy from 70.3 percent to 4.6 percent at a constant 1 percent false positive rate, while also evading watermarking, GPTZero and OpenAI's classifier (Krishna et al. 2023, arXiv 2303.13408, read 9 November 2026). Sadasivan and colleagues showed that recursive paraphrasing of passages around 300 tokens long significantly reduced detection rates across watermark, neural, zero-shot and retrieval detectors while only slightly degrading text quality in many cases (Sadasivan et al. 2023, arXiv 2303.11156, read 9 November 2026).

Commercial humanizers are that finding, productised. None of the twelve tools in this site's set publishes its model, so the only way to know what a given tool does to a given text is to run it, which is what the published method does.

People mix these up because several vendors sell all four. The differences are in what each one promises and measures.

The humanizer and the paraphraser are the same machine with different marketing. QuillBot's humanizer caps free input at 125 words and offers Basic and Advanced modes, checked on 8 November 2026; its paraphraser has lived on the same site for years. Grammarly prices its plans in "AI prompts" per month rather than humanizer words, also checked on 8 November 2026, which tells you the humanizer is a feature of a larger product rather than the product.

Detectors and humanizers are now locked together. A detector maker trains on humanizer output; a humanizer vendor tunes against detectors. Turnitin states that its indicator covers "AI-generated text that may have been modified by AI paraphraser or bypasser (also called humanizers) tools" and that its paraphrase detection is available only for English submissions (Turnitin, AI writing detection FAQ, read 9 November 2026). A humanizer's result is therefore true for one detector on one date, and the guide on whether AI humanizers work walks through the evidence on both sides.

The sequence runs from detector to study to product. OpenAI released a classifier for AI-written text in January 2023 that identified 26 percent of AI text as "likely AI-written" and mislabelled human text 9 percent of the time, and retired it on 20 July 2023 "due to its low rate of accuracy" (OpenAI, new AI classifier announcement, read 9 November 2026). The same announcement listed as a known limitation that "AI-written text can be edited to evade the classifier".

The March 2023 papers from Krishna and Sadasivan turned that limitation into a measured attack. Within months, paraphrasing services began selling themselves as humanizers, with the detector readings as the pitch. Three years on, the market has three recognisable shapes:

• Standalone humanizers with a strength ladder and a per-request cap: Undetectable AI, StealthGPT, Phrasly, BypassGPT, HIX Bypass, WriteHuman, GPTHuman, Walter Writes AI, Verva (formerly TwainGPT).
• Humanizers bolted onto a writing suite: QuillBot and Grammarly.
• Free tools with no account: Clever AI Humanizer, free with a 3,000-word request cap and no sign-up, checked on 8 November 2026.

The names also shift. TwainGPT became Verva, and HIX Bypass moved to hixbypass.com, both noted on 8 November 2026.

The business model is words. Every paid plan sets words per month and most set words per request, and the request cap is the number that bites, because it decides how many pieces a long document is cut into.

All figures were read on each vendor's pricing page on 8 November 2026 and are listed per tool, with the same fields, in the reviews. Price is reported on this site and never scored; the method's limits-and-access dimension scores words per request, words per month, free allowance and sign-up, not dollars.

It does not check your facts; a paraphraser can drop a qualifier or swap a technical term and nothing in the interface will tell you. It does not fix grammar, and stronger modes add errors. It does not give you a Turnitin result, because Turnitin's detector is not available to the public and no vendor can test against it. And it does not change what your institution's or employer's AI policy says.

No detector score should be the only evidence in a misconduct case, and no humanizer output should go anywhere without the writer reading it against the original.

Four fields, the same for every tool: did the meaning survive, how many errors were added, how did named detectors read it on a stated date, and what did the plan allow. That is the rubric in the published method, weighted 35, 20, 30 and 15. A vendor claim that answers none of those is marketing; a tool page in the reviews answers all four with dates. The guides hub holds the rest of this series, including the difference between a humanizer and a paraphraser in more detail.