Should I use an AI humanizer or rewrite myself?

The last row is the one that decides most academic cases. Stanford's default policy, adopted on 15 October 2026, permits "surface-level editing that does not materially alter the substance, analysis, organization, or argument" and forbids AI tools that "create deliverables" (Stanford Office of Community Standards, generative AI policy guidance, read 9 November 2026). A hand edit of your own permitted draft sits inside that line; a wholesale AI rewrite does not.

The same things, by different hands. The University of Toronto's FAQ describes what detectors measure: "sentence complexity, vocabulary richness, and syntactic variety", and it states that writing that is "too simple or too uniform, or unusually polished with a lack of variation" may be flagged (University of Toronto, generative AI FAQs, updated October 2026, read 9 November 2026). Both routes act on exactly those properties.

None of the vendor pages checked on 8 November 2026 states an error rate, a meaning-retention figure or a processing time. They state word caps, prices and mode names. The guide on how to humanize AI text manually lists the edits a writer makes by hand; the guide on how to humanize AI text covers the tool route.

One peer-reviewed study tested both routes on the same inputs. Weber-Wulff and colleagues prepared 54 documents in 2023, including AI-generated texts that a human had edited "exchanging some words with synonyms or reordering sentence parts" and AI-generated texts rewritten by QuillBot on its default Standard mode, and ran all 54 through 14 detectors, 756 tests in all (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026).

Two things follow. First, both routes moved detector readings on 2023 models, and the tool route moved them further; the authors called the paraphrase result the most surprising in the paper and concluded that about half of all obfuscated AI text would be misattributed to humans. Second, the study measured only detector readings. It did not measure whether the edited or paraphrased texts still said what the originals said, how long each route took, or how many errors each introduced. No detector score should be the only evidence in a misconduct case, and no detector score tells a writer whether the rewrite is still true.

Those two figures do not yet exist in any source this page could cite, from a vendor or a study. The vendor pages state caps and prices. The study states detector accuracies. Neither states minutes per 500 words or errors per 100 words for a hand rewrite against a tool rewrite on the same text.

The published method on this site is built to produce both. Each run takes 24 AI-written inputs across six genres, runs every tool on its default and strongest mode, and has two editors rate meaning kept on a 1 to 5 scale and count errors introduced per 100 words, with meaning weighted at 35 of 100 points, errors at 20, detector pass at 30 and limits at 15. The first run will add a hand-rewrite row to the same 24 inputs, timed per input and read by the same two editors, so the time and error columns for both routes appear on the runs page with the raw files when it publishes.

Until then, the table above is the evidence: a humanizer's output arrives in seconds and in pieces no longer than its request cap, and a hand rewrite arrives at the speed of your reading, with your meaning intact by default.

• Choose to rewrite yourself if the text is under about 500 words, you know the subject well enough to catch a wrong term, and the writing is meant to sound like you.
• Choose to rewrite yourself if the rule you work under permits editing your own permitted draft but bans AI tools, as Stanford's default policy does.
• Choose a humanizer if you have more text than you can retype, the use is permitted and disclosed, and you will read every output line by line against the input before anything else.
• Choose a humanizer if you want a mild mode as a first pass and intend to finish the edit yourself.
• Choose neither if the question is really whether an AI draft was allowed; settle that first.

The rest of this series sits in the guides hub.