Try the September 2026 engine on your own draft
Free account: four runs of 250 words, once. Score the result in the next tab.
Five detectors, one essay, five human verdicts. On 23 September 2026 we took a single HumanizeMyAI rewrite, a 353-word academic essay with APA-style citations, and pasted it into each detector's own interface. Nothing was edited between runs.
| Detector | Result, 23 Sep 2026 |
|---|---|
| GPTZero (model 4.10b) | Human 100%, AI 0% |
| Pangram | 100% human-written |
| Copyleaks | 0% AI, no AI content found |
| Grammarly AI Detector | 0% AI |
| Originality AI | Likely original (15% or less, the finest reading its free check gives) |
Not one of the five flagged the essay, and the two with a single headline number, GPTZero and Pangram, both landed at the top of the human end of the scale. What Pangram's labels mean, tier by tier, is set out in our Pangram detector review.
Five detectors, five human verdicts
Every row in the table comes from a screenshot taken during the run. We cropped each one to the verdict so the number is readable and nothing else from the session shows.
Cited academic prose is the hardest case for a humanizer. Author names, years and technical terms cannot be rewritten without breaking the paper, so every one of them has to survive untouched while the sentences around them change. A rewrite that clears detectors by deleting references is useless to a student. The September essay kept its citations, and the verdicts above are for that version.
The 31 August run used a different text, a 668-word academic essay, and a different detector set. Taken together, the two runs cover eight detectors and two unrelated essays.
| Detector | Result, 31 Aug 2026 |
|---|---|
| GPTZero | 0% AI |
| Turnitin | Human (no score shown under 20%) |
| Originality AI | Human (15% or less) |
| Copyleaks | 0% AI |
| QuillBot AI Detector | 0% AI |
| ZeroGPT | 0 to 3% AI |
GPTZero, Copyleaks and Originality AI appear in both runs and gave the same human reading a month apart, on different essays. That repeat is the part we care about: a result that holds across texts and dates is a property of the engine, not of one lucky paragraph.
The engine behind these runs is the one live on the site today. No beta, no special mode, no waiting list. Anyone who pastes text into the humanizer on our homepage gets the same build that produced the screenshots above.
Two smaller changes shipped with it:
The September engine replaced the build released on 13 May 2026, and that release made four changes. Sentences and paragraphs were rebuilt instead of swapped word for word. A cleanup pass removed leftover formal academic phrasing. The reference examples were refreshed from the 2,590-essay corpus. And testing on subjects the engine was never tuned on, such as psychology, art history, environmental economics and education theory, became part of every release.
In that May test, 30 of 30 humanized passages came back clean on QuillBot's own AI detector, up from 16 of 30 on the build before it. Those May figures are a dated record now. The two tables above, from 31 August and 23 September, describe the engine live on the site today.
Run your draft through our free AI detector first, then humanize it and scan it again in whichever detector your school or client uses. Detector vendors update their models, so the right habit is to check the current version before anything important goes in, the same way we date every row on this page.