Yes, in most cases. And when a humanizer does manage to lower Compilatio's AI detection score, it almost always comes at the cost of a degraded text, bordering on incoherence.
We explain what these tools are, what our first internal tests show, and where we will keep digging.
1. What is a humanizer tool?
A humanizer tool (or “humanizer”) is software that rewords an AI-generated text in an attempt to hide its origin and make it look as if a human wrote it. In practice: you paste a text produced by ChatGPT, Gemini or another model, the tool changes its words, sentence structure and rhythm, and returns a version that is supposed to slip under the radar of detectors.
Not all of these tools were originally built for that purpose. Some, like Quillbot or Wordtune, started out as rewording and writing assistance tools. Others, like Undetectable AI or BypassGPT, were designed from the start to get around detection, and say so openly in their name and marketing promise.
As schools and universities have equipped themselves with AI detectors, a parallel market of tools to get around those same detectors has grown up in mirror image, all over the world. The two markets have moved forward together since 2023: the more widespread the detection of AI-generated texts becomes, the more sophisticated the bypass tools get, and vice versa.
2. How Compilatio holds up against bypass tools
We ran an internal assessment to measure how well our detection really resists bypass tools, using the 6 most widely used humanizer tools in the world.
The methodology:
300 texts were tested, generated by two AIs (ChatGPT and Gemini), reworded by the 6 most popular humanizer tools in the world, in French and in English. For each text, we observed how the Compilatio detection score changed before and after the text went through the humanizer, and we qualitatively assessed the resulting rewording.
Summary of results:
The tools tested are deliberately anonymized. The goal of this article is to explain the underlying trend we observe in our detection tool.
| Tool | Change in Compilatio detection score before and after humanization | Text quality* |
| Humanizer 1 | Sharp drop | 1.4 / 5 |
| Humanizer 2 | Sharp drop | 1.5 / 5 |
| Humanizer 3 | Moderate drop | 1.8 / 5 |
| Humanizer 4 | Slight drop | 2.3 / 5 |
| Humanizer 5 | Up | 3.6 / 5 |
| Humanizer 6 | Up | 3.8 / 5 |
*Internal qualitative assessment out of 5, not an automated measurement: to be taken as a trend, not as an absolute score.
What we observe:
- The tools that lower the detection score the most are also the ones that degrade text quality the most, to the point of making it unusable in an academic setting.
- The tools that best preserve text quality do not lower the detection score; on the contrary, they even push it up.
The conclusion fits in one sentence: to date, none of the 6 tools tested manages both to effectively fool detection and to produce a text of good academic quality.
→ Either you drastically degrade the quality of your text in the hope of going unnoticed.
→ Or you keep a quality text, and the AI is still detected, often even more than before.
In both cases, the time spent getting around detection would have been better invested in building expertise on the subject at hand.
3. What's next
This topic moves fast, and so do the tools. Here is what we will be working on very soon:
- Expand our test corpus, with more texts and more humanizer tools, to refine and consolidate these first observations.
- Expand the number of languages tested, continuing with French and English, and adding Spanish and Italian.
- Keep training our detection model, on both standard AI generation and humanization techniques.
This article has been automatically translated. If you notice a translation error, please contact us.