Does an AI humanizer fix grammar, or add mistakes?

Three kinds of statement appear on the vendor pages, and none is "we correct your grammar". The first kind describes what the rewrite changes. Grammarly's AI Humanizer page says the tool scans for "an overly formal tone or repetitive phrasing" and refines text "to make it sound more natural and engaging while preserving accuracy", and states that it "is not intended to bypass AI detectors" (Grammarly, AI Humanizer, read 9 November 2026). Tone and phrasing, not errors.

The second kind admits the output may need a proofreader. GPTinf's home page FAQ, answering "what is the quality of the final result", says meaning, style and semantic structure are preserved and adds: "We recommend using Grammarly after paraphrasing to fix possible grammatical errors (it does not affect the detection score)", read 9 November 2026. A vendor telling buyers to run a second tool on its output is the clearest statement on this page.

The third kind is a feature label. HIX Bypass's pricing page lists "Error-Free Outputs" among the features of every plan, beside "Plagiarism-Free Content" and "No Distortion Humanization" (HIX Bypass, pricing, read 9 November 2026). The page gives no error rate, no test and no definition of error. GPTHuman's home page takes the same position from the other side, telling buyers to "avoid using tools that will add grammar and syntax errors to your content", read 9 November 2026, without a figure for its own output.

No, and Grammarly's own pages draw the line. The Grammar Checker page describes a tool that finds "grammatical mistakes, spelling errors, and typos" and lists its free tier as grammar, spelling, punctuation, conciseness and tone detection, with full-sentence rewrites, word choice and fluency suggestions on Pro (Grammarly, Free Grammar Checker, read 9 November 2026). The AI Humanizer page describes a different job: rewriting AI text so it "sounds natural and engaging while staying true to your original message", with four preset voices and six languages, read 9 November 2026.

The distinction matters for a buyer who assumes the humanizer inherits the checker. Grammarly prices Pro in "AI prompts" a month rather than words, read 9 November 2026, and the humanizer page says nothing about the checker running on its output. The Grammarly AI Humanizer review treats the two products separately for that reason.

The one peer-reviewed study that tested edited and paraphrased AI text measured detector readings, not error counts, and its findings still bear on this question. Weber-Wulff and colleagues ran 54 documents through 14 detectors in 2023 and found that unmodified AI text was correctly classified 74 percent of the time, AI text lightly edited by a human 42 percent, and AI text paraphrased by a tool 26 percent (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). They described the human edits as "exchanging some words with synonyms or reordering sentence parts".

Synonym swaps and reordered clauses are what move a detector reading. They are also the two operations that produce the writing faults this site counts: a synonym misfire, where the replacement word is not a synonym in context; a replaced technical term; a fragment or run-on from a clause moved without its punctuation; a pronoun left without a referent. The study did not count those faults. It did show that the edits detectors respond to are the edits grammar is most exposed to, and GPTinf's statement that a grammar pass "does not affect the detection score" is the same point from the vendor side: corrections do not move readings, rewrites do.

For every output, two editors reading blind to the tool will count errors introduced per 100 words under six classes, and that figure carries 20 of the 100 points on the method page. The six classes are agreement and tense faults, synonym misfires, replaced technical terms, fragments and run-ons, broken references, and broken punctuation or list structure. Each marked error is recorded with the sentence it sits in, so a vendor can check the count against its own output. The classes are illustrated, with before and after pairs written by the editors, on the humanizer errors page.

Until the first run publishes, the practical advice is the order of operations. Read the output against the input for meaning first, because a grammar checker cannot see a dropped claim or a qualifier that changed from "may" to "will". Then run the grammar checker, which is the step GPTinf itself recommends. No detector score should be the only evidence in a misconduct case, and no clean grammar check is evidence that a rewrite still says what you meant. The rest of this series is in the guides hub.