An AI Humanizer Rewrites Text to Dodge Detectors and Tends to Damage the Prose

It rewrites the sentence, one word at a time in the mildest version, and hopes the result reads as a person. Pangram Labs, the detector company, named three techniques in a post dated 27 January 2025. Humanizers swap words for their synonyms. They add nonsensical phrases, on the theory that a detector reads gibberish as unlikely to have come from a machine. The third technique is plain damage to the text quality ¹.

The synonym pass is the quiet one. Clause order holds, sentence length holds, and the content words arrive slightly wrong for the register. The same company's January 2025 research paper audited 19 humanizer and paraphrasing tools and found that split running through the whole category: some tools retain low-level sentence structure and replace individual words with synonyms, while others take more liberty, rewording entire groups of sentences and paragraphs, adding sentences that were not in the original or deleting redundant ones ².

The inserted noise is far easier to see than the synonyms. That audit collected fabricated references and strings of question marks left mid-paragraph. One passage trailed off into the fragment "CGSizeMake pp 18-23" where a citation belonged ².

Vocabulary level is the tool's decision rather than the paper's. Some humanizers write only in a formal university register. Others drop to elementary or high school level, and the better ones, usually those built on language models, adopt the tone of the document handed to them ².

What that leaves on the page has its own signature, and a teacher can learn it by eye: what humanized text looks like.

The rewrite costs quality, on the one measure the auditors ran and in the tiers they drew by hand. The 2025 study asked GPT-4o to pick the more fluent and coherent of two passages, the original or the humanized version, over 25 samples per tool. The best tier of humanizers won 26.0 percent of those comparisons, the middle tier 14.67 percent, and the worst tier 2.67 percent ².

Read the top figure the right way round. A win rate of 26.0 percent means that about three times in four, GPT-4o picked the original as the more fluent and coherent passage ². The paper's own summary is flat: all humanizers tend to degrade the quality of the original text, though the degree varies ².

Its three tiers were set on faithfulness and fluency alone, not on whether a tool beat a detector. At the top, tone, vocabulary level and complexity survived. The middle tier degraded the writing but kept the intent. At the bottom, nonsensical phrases often went in and sentences became uninterpretable. Meaning often shifted with them ².

A paragraph from that worst tier, carrying an invented reference, a broken subject and verb, and a key term traded for a near-synonym that means something else, fails on its own merits before any detector runs on it.

None of that damage touches the integrity question. At the University of Southern California, the academic integrity office states that material created by a generative tool and presented as a student's own work counts as plagiarism whether it was copied verbatim, near-verbatim or paraphrased ⁶.

Under that wording a humanizer pass is paraphrase, and changes nothing. Other institutions write their own rule, and local policy governs: a humanizer is not a plagiarism remover.

Sometimes, and the published results disagree. In Pangram Labs' own January 2025 test, its detector trained with humanized examples held 98.26 percent on humanized academic text ²; an independent 2026 study found fewer than 4 percent still flagged ³. The two figures come from different test sets, different detectors and different years, so neither settles the other.

In that same January 2025 test, at a 5 percent false positive rate, GPTZero as tested then fell from 99.73 percent on machine-written passages to 60.04 percent once a humanizer had passed over them, and Binoculars from 94.15 percent to 28.23 percent ².

Pangram's July 2026 technical report puts humanized text at 97.67 percent labelled AI-generated, with a per-system true positive rate running from 91.52 percent to 99.39 percent across the commercial tools it evaluated ⁵. All of these are figures from the company selling the detector, including the ones about its competitors.

The measurement pointing the other way is the 2026 University of Notre Dame study. It built machine rewrites of 642 published abstracts of 25 to 500 words, then ran the rewrites through a commercial humanizer. Fewer than 4 percent of the rewrites a detector had flagged as AI were still flagged afterwards ³. A separate Pangram test, reported by Nature, was of human-written student essays substantially modified with consumer AI, not of humanized machine text; 41 percent of those results were still labelled fully human ⁴.

Nature's caution about the whole category is the sentence to keep: only the accuracy rates the companies announce from their own internal testing are current, and those are not externally verified ⁴. Whether these tools work, and for how long, is its own question: do AI humanizers work.

Each technique lands in a different part of the paragraph. The synonym pass alters vocabulary and nothing else. Insertion adds material that was never in the draft. Degradation surfaces as misspellings and grammatical errors ¹², which on their own look like any hurried draft and are not evidence of a tool.

Price is a separate axis, and Human, the detector at human.olive.is, reports one finding on it. Whether a humanized document is caught depends on the humanizer's price tier more than on which humanizer it is. That reading is Human's own in-house measurement on its own held-out sets as of September 2026, not an independent audit. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding.

The student side of that transaction is free humanizer versus paid tier.

The market behind these tools is neither small nor hidden. NBC News reported that Turnitin keeps a list of 150 such tools, some charging as much as 50 dollars for a subscription ⁷. An academic integrity company counted 43 humanizer sites drawing 33.9 million visits in a single month ⁷.

Grammarly's AI detector guide says the company does not currently offer a dedicated feature to reduce an AI score ⁸. Its ordinary red and blue underline corrections should not normally move a detection percentage, though generative rewriting of whole sentences and paragraphs will raise one ⁸. Grammarly also publishes a separate AI humanizer, and the page for that tool states that it is not intended to bypass AI detectors ⁹. Where that line falls inside a school is Grammarly's humanizer in school.

Run the full submission rather than the three sentences that read oddly, and write the tool and version beside any score that goes into a file. The Notre Dame cohort was abstracts of 25 to 500 words ³, so its figures describe short passages and may not carry to a full paper.

Then ask what a rewrite cannot answer. The draft history is quick to ask for and settles a question a second detector run cannot: Google Docs version history. A short conversation about the argument is the other, and it works whether or not any tool was involved: an oral follow-up after a flag.

One more finding from the Notre Dame paper belongs in the same conversation. Disclosed, honest AI editing carried a higher risk of sanction than concealing machine drafting behind a humanizer ³. The paper that declared the help reads as machine-edited. The paper that hid it behind a humanizer reads as clean. That asymmetry is an argument about policy, not about any one paper.

Length decides what a reading is worth, whichever tool produces it. At human.olive.is, Human reads college-level academic writing, reports sentence by sentence how much of a document reads as machine-written or machine-edited, and returns a document verdict of Human, Mixed or AI. The public check reads up to 1,000 words, three times a day, with nothing stored and no account, so a longer submission does not fit in one paste. Under 50 words it declines to give a reading at all. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding.