Which AI humanizers pass AI detectors
One table, six detectors, the same three test texts for every tool, and a date on every row. A detector counts as passed at or below 35% AI probability. Vendor claims are on the individual detector pages, kept well away from this one.
This table is empty, and that is the honest state
Every row here has to come from a person running three fixed test texts through a humanizer and then through six detectors, and writing down what happened. None of those runs has been published yet, so there is nothing in the table.
The alternative - filling it with the "99.8% undetectable" numbers the vendors publish - would make this page look finished today and worthless by the time anyone relied on it. Those claims are listed, clearly labelled as claims, on each detector's page.
In the meantime the method is published in full, and it is the same method whether we run it or you do: the test procedure.
What a filled row will look like
An illustration of the format, not a result. No tool is named because no tool has been measured.
| Detector | Before | After | Result |
|---|---|---|---|
| GPTZero | 98% | 22% | Improved |
| Originality.ai | 94% | 41% | Partial |
| ZeroGPT | 99% | 11% | Strong |
| Copyleaks | 91% | 68% | Weak |
One detector at a time
Turnitin has a page but no numbers: it is institution-controlled, there is no public interface to submit text to, and anyone showing you Turnitin percentages from a consumer tool is showing you something else.
What a passing result is not
It is not a guarantee, and it is not permission. It is one measurement, of one piece of text, by one detector, on one day. Detectors are retrained without notice; a result from September can be wrong in December, which is why the date is part of every row.
Whether you may use any of these tools is decided by your institution, employer or client. See academic integrity.