Does humanizing AI text fool AI detectors? Test results
Partly. We rewrote Claude Opus 5.5 texts with the open-source humanizer skill, which removes common AI phrasing without changing the content. Desklib's catch rate fell from 62.7% to 32.5%, Fast-DetectGPT's from 63.5% to 47.6% and MAGE's from 45.2% to 23.8%. Many humanized texts were still flagged.
Tested
How much does editing change detection?
| Detector | Original | Humanized |
|---|---|---|
| Fast-DetectGPT (Qwen2.5-1.5B) | 63.5% | 47.6% |
| Desklib AI text detector | 62.7% | 32.5% |
| Fakespot AI text detector | 65.1% | 47.6% |
| MAGE | 45.2% | 23.8% |
| E5-small LoRA detector | 64.3% | 47.6% |
| OpenAI RoBERTa detector | 2.4% | 0.0% |
| TMR (Target Mining RoBERTa) | 83.3% | 81.0% |
What we did
We took the same Claude Opus 5.5 texts used in the rest of the test and rewrote each one with the humanizer skill, run by Claude Opus 5.5 itself. The skill removes patterns such as stock phrases, forced lists of three, dashes and empty closing lines, and it may not add or remove facts. The rewrites stayed close to the originals in length and content. We then scored the rewritten versions with all seven detectors.
The humanizer’s own documentation says it edits for human readers and that detectors still flag most of its output. That did not hold for these detectors: six of the seven flagged fewer than half of the humanized texts, and the seventh, TMR, flags most human writing too.
Results
Desklib was hit hardest: its catch rate fell from 62.7% to 32.5%. MAGE dropped from 45.2% to 23.8%. Fast-DetectGPT, Fakespot and E5-small LoRA all fell to 47.6%, from 63.5%, 65.1% and 64.3%.
TMR barely moved, from 83.3% to 81.0%, because it flags most text whoever wrote it. The OpenAI RoBERTa detector flagged none of the humanized texts, but it caught only 2.4% of the originals.
Reading a detector result on edited text
Light editing removes much of what these detectors look for. A detector that catches about six in ten raw Opus texts, as Desklib and Fast-DetectGPT do, catches roughly a third to a half after one editing pass. A “human” verdict on edited text is weak evidence of anything.
Questions people ask
What is an AI humanizer?
A tool or prompt that rewrites AI-generated text so it reads more like a person wrote it. We used the open-source humanizer skill by blader, which works from Wikipedia's guide to the signs of AI writing.
Do AI humanizers make text undetectable?
Not in our test. The humanized texts were caught less often, but Fast-DetectGPT still caught 47.6% and Fakespot 47.6%.
Why test humanized text at all?
Because much AI-assisted writing gets edited before anyone checks it. A detector result on raw chatbot output tells you little about text that a person or a tool has rewritten.