Desloppy vs six AI detectors on humanized text
A public dataset scores detectors on AI text that was run through "humanizer" tools. We scored the same texts with Desloppy. Here is what came out, with the caveat that matters.
People who want AI text to get past a detector run it through a "humanizer". These tools rewrite the draft to hide the chatbot sound. A detector that misses humanized text is easy to get around. So we tested Desloppy on exactly that kind of text, and we compared it with six other detectors on the same texts.
The texts
The texts come from a public dataset called ai-humanizer-benchmark, published under a CC-BY-4.0 licence. We used the cycles from September and October 2026. Together they hold 726 AI texts rewritten by commercial humanizer tools. The dataset also publishes the score each detector gave every text. That is what makes the comparison fair: every detector saw the same texts.
Desloppy had already trained on 508 of those texts, so we threw those out. The fair subset is the 218 texts Desloppy never trained on. Most of them come from three tools we hold out from training for exactly this purpose: 66 from hix-bypass, 66 from stealthgpt and 65 from undetectable-ai. The other 21 come from other tools. The average text is 336 words long.
How we scored
We ran the live Desloppy model, version v20261018-1313-c2, on our own machine on 20 October 2026. It gives the blended Slop score, the same number you get on the site. A text counts as caught when its Slop score is 50 or more.
For the other detectors we used the scores the dataset publishes. Each of them reports how likely it thinks a person wrote the text. A text counts as caught when that number is under 0.5.
The results
Out of the 218 humanized texts, this is the share each detector caught.
| Detector | Caught |
|---|---|
| Desloppy (Slop score) | 72.0% |
| GPTZero | 68.8% |
| Desloppy (model alone) | 63.8% |
| Originality.ai | 60.6% |
| Winston AI | 51.4% |
| QuillBot | 39.0% |
| Grammarly | 34.9% |
| ZeroGPT | 33.5% |
Desloppy caught 72.0% of the texts. GPTZero caught 68.8%. That gap is 3.2 points, and the 95% range runs from 5.0 points behind to 11.5 points ahead. The statistical test gives p = 0.52. In plain words, Desloppy and GPTZero tied on this set.
The other gaps are clear. Desloppy caught 11.5 points more than Originality.ai, with a range of 5.0 to 17.9 points and p = 0.001. It caught 20.6 points more than Winston AI (range 12.4 to 29.4), 33.0 points more than QuillBot (range 24.8 to 40.8), 37.2 points more than Grammarly (range 29.8 to 44.5) and 38.5 points more than ZeroGPT (range 30.7 to 46.3). Each of those four has p under 0.0001.
The "model alone" row deserves a look. Desloppy's detection model on its own caught 63.8%. The Slop score blends that model's reading with the sloppy lines Desloppy finds in the text, and the blend caught 72.0%. The sloppy lines do real work here. Humanizer tools change words, and the stock phrases often survive. The blend is explained in What the Slop score is.
The dataset also scores a seventh detector. We leave it out of this write-up.
What this shows, and what it leaves open
On this set, Desloppy caught more humanized AI text than Originality.ai, Winston AI, QuillBot, Grammarly and ZeroGPT, and tied with GPTZero. That is the whole claim.
It says nothing about overall accuracy, and it allows no plain "better than" without that qualifier. This test measures one side only: AI text caught. The other side is how often each detector wrongly flags writing a person did by hand, on the same kind of texts. A detector that flags everything would score 100% here and be useless. We publish our own wrong-flag rates on the About page, but the dataset has no matching numbers for the other tools, so the two sides can't be compared yet.
There is a second reason for care. Each vendor sets its own threshold. A stricter threshold catches more AI text and flags more hand-written text. A looser one does the opposite. Moving the line moves both numbers at once.
Pangram is missing because the dataset has no scores for it.
The next test
The test we want to run next measures both sides. It would use about 500 texts held out from training: AI text from current models plus licensed human writing, short and long, in everyday genres. Desloppy, GPTZero, Originality.ai and Pangram would score every text at each vendor's default setting. We would report AI text caught, hand-written text wrongly flagged, and AUROC, a measure of how well a detector ranks texts across every threshold. That test needs paid vendor accounts, so it waits for a budget.
If you check your own writing, remember what the score is. The Slop score is a probability. It says how much a text sounds like default chatbot writing, and nothing about who wrote it. You can try 1,000 words free, with no account, at desloppy.com.