How to humanize AI text

Humanizing is only useful if you know what you are changing. Three kinds of signal cover most published detectors.

Liang and colleagues found that the TOEFL essays all seven detectors in their study flagged as AI had significantly lower perplexity than the rest, and that a prompt asking for richer language cut the average false positive rate on those essays from 61.22 percent to 11.77 percent (Liang et al. 2023, Patterns00130-7), read 9 November 2026). GPTZero describes its current model as a sentence-by-sentence classifier with a "Paraphraser Shield" against paraphrasing and homoglyph attacks (GPTZero, technology, read 9 November 2026). Turnitin states that false positives can include "content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas" (Turnitin, AI writing detection FAQ, read 9 November 2026).

Put plainly: detectors read predictable words, uniform structure and the absence of anything a person would have added. The guide on what AI detectors measure goes deeper; the edits below follow from those three signals.

Start with the edits that add information, because they are the ones a paraphraser cannot make for you.

The last two rows are where Weber-Wulff and colleagues' 42 percent figure came from. In their study, the manually edited case involved a person "exchanging some words with synonyms or reordering sentence parts", and that alone cut detection from 74 percent to 42 percent across the 14 tools (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). The authors called this patchwriting and noted it is the most common form of misuse, which is exactly why the detector makers have spent the years since training against it.

A reader who wants the whole checklist should use the 20-edit manual guide on this site, which works through each edit with an example. The order matters: facts first, structure second, word swaps last, because the word swaps carry the most meaning risk for the least gain.

Three situations. The text is long enough that hand editing takes hours. You have already made the information edits and want a second pass on word choice. Or you are testing several versions against a detector and need speed. In the 2023 study, the machine paraphrase case left only 26 percent of AI texts detected, the lowest of any case, so the tools do move readings; the question is what they cost.

All five limits were read on each vendor's own pricing page on 8 November 2026. Walter Writes AI names three modes, Simple, Standard and Enhanced, and its 750-word request cap means a 2,000-word essay goes through in three pieces, each of which can lose context at the seams. The other eleven tools in our ranking page have their own caps and modes, and the ranking lists them on the same fields.

Whatever the tool, run it on text you have already edited for facts and structure. A paraphraser cannot add a detail you did not give it, and the detail is the part that reads as human.

The published studies measured detectors, not meaning. Weber-Wulff and colleagues documented the differences between original and obfuscated texts in an appendix but scored only detection. Krishna and colleagues described their paraphraser as not appreciably modifying the input semantics, by automatic measures. Nobody in the 2023 literature had an editor read the output against the input and count what went missing.

That gap is the reason this site's published method weights meaning kept at 35 points and errors added at 20, against 30 for detector pass rate. The things to look for after any humanizing pass, by hand or by tool:

• A qualifier dropped: "in most trials" becomes "in trials", and the claim is now broader than its evidence.
• A term swapped: "statistically significant" becomes "important", and the sentence now says something different.
• A sentence that no longer parses, usually where two clauses were merged.
• A new claim: the smoothest rewrites sometimes assert a cause the input only implied.

No detector score should be the only evidence in a misconduct case. The same rule applies in reverse: no detector score should be the only check on a rewrite. Read the output against the input, line by line, before you use it.

Run the finished text through at least two detectors, not one, and note the date. Detectors disagree: in the 2023 study, false positive rates on human writing ranged from 0 percent for Turnitin to 50 percent for GPTZero on the same 54 documents, so a pass on one tool tells you little about the next. GPTZero states that its confidence bands are tuned so the average error rate is under 1 percent for its high-confidence predictions, and that figure is from its own internal evaluation set (GPTZero, technology page, read 9 November 2026).

Then read the text aloud. If a sentence sounds like something you would not say, that is the sentence a classifier was trained on. Fix it with a fact, not a synonym. For the broader question of whether any of this holds up over time, the guide on whether AI humanizers work sets out what the studies found and what detector makers now publish about humanizers; the guides hub holds the rest of the series.