
Short answer: the Scribbr AI detector is careful with human writing and weak on reworded AI text. Scribbr's own 30-text test scored the checker on its detector page 78%, but a 2026 Lublin University of Technology study scored it 44.0%, and it flagged none of 60 AI-paraphrased samples. The same Lublin study scored ZeroGPT at 64.0% and GPTZero at 58.7%, a gap our ZeroGPT vs GPTZero guide unpacks. Treat a score as a hint.
I read Scribbr's detector page and its 12-tool test on 16 November 2026, then 2 independent studies. One September 2026 review says "no independent researcher has published a test of Scribbr at all." There are 2, and their numbers look very different.
How accurate is the Scribbr AI detector in independent tests?
Less accurate than Scribbr's own figure suggests. In a 2026 test of 150 Polish and English samples, the Scribbr AI detector scored 44.0% overall accuracy against 64.0% for ZeroGPT. It never flagged human text as AI, but it missed most AI text, including all 60 AI-paraphrased samples.
| Test | Texts | Human texts flagged as AI | AI text caught |
|---|---|---|---|
| Scribbr's own test (published April 2026) | 30 | 0 of 5 | 78% overall score |
| Brodacki and Plechawska-Wójcik, 2026 | 150 | 0 of 30 | 36 of 60 fully AI, 0 of 60 paraphrased |
| Lege, 2025 | 912 | 20 of 605 at a 30% cut-off | 307 of 307 |
Marcin Brodacki and Małgorzata Plechawska-Wójcik ran Scribbr's web checker from 6 to 9 May 2026. Scribbr and Copyleaks came out statistically level, behind GPTZero and ZeroGPT.
R. Paul Lege at Nagoya University ran the Scribbr AI detector on 307 AI texts, 302 published human texts and 303 theses by second-language graduate students. It's a single-author audit, so weigh it lightly.
What does Scribbr's own 78% figure actually measure?
A small vendor test. Scribbr ran 30 texts, 5 in each of 6 categories, each 1,000 to 1,500 characters long, through 12 detectors. The checker on its AI detector page scored 78%, and a second Scribbr version topped the table at 84%.
Koen Driessen's 12 Best AI Detectors article (published 22 April 2026, revised 21 July 2026) scored each text 1, 0.5 or 0, yet its results table is still headed "Best AI detectors in 2024". It says the 78% checker "did not incorrectly label any human-written text as AI-generated".
Two things bugged me. Scribbr and QuillBot, which tied for second, are "sister companies", in Scribbr's own words. And Scribbr's detector page says the best result it found for a checker like that one was 68%, while the linked test puts that very checker at 78%.
Does Scribbr flag human writing as AI?
Rarely, in tests. Scribbr's own run flagged 0 of 5 human texts and the 2026 Polish study flagged 0 of 30. Lege's audit, at a 30% cut-off, flagged 1 of 302 published texts and 19 of 303 theses by second-language writers: 6.3% of the theses against 0.3% of published texts.
Count any score above zero and the picture changes: the Scribbr AI detector put some AI percentage on 211 of the 303 theses, including those 19, and on 101 of the 302 published texts. Averages were 15.4 against 5.9. Lege writes that AI detectors "disproportionally flagged ESL writing with false positives."
So a 15% reading on a second-language writer's essay proves nothing. Our guide to Turnitin false positives shows the same problem elsewhere.
Can the Scribbr AI detector catch reworded AI text?
Mostly no. In the 2026 Polish study, the Scribbr AI detector labelled 28 of 30 Polish paraphrased samples human and the other 2 uncertain, and called all 30 English paraphrased samples human. Scribbr's own test caught 50% of its mixed and paraphrased texts. Coin toss.
Fully AI text wasn't a sure thing either: 22 of 30 Polish AI samples caught, but only 14 of 30 English ones. Turnitin has a similar weakness, as our breakdown of whether Turnitin can detect AI humanizers shows.
What do Scribbr's 4 result labels mean?
Scribbr sorts each section into 4 groups: AI-generated, AI-generated and AI-refined, human-written and AI-refined, or human-written. The overall score runs from 0% to 100% and shows "how much of the text contains content likely written or refined using AI tools".
A grammar-polished human essay can land in a refined group, and none of the 3 tests checked those labels on their own. Our Grammarly AI detector review covers how editing tools move scores.
Does a low score predict your Turnitin report?
No. They're separate tools from separate companies, and none of the 3 tests above included Turnitin. A 0% from the Scribbr AI detector tells you nothing certain about your university's report. I searched Scribbr's detector page for Turnitin and found no mention; it sends readers to QuillBot's detector for a second check.
Our QuillBot AI detector review and GPTZero vs Turnitin comparison show how far two detectors drift.
How can you test the Scribbr AI checker yourself?
Copy Scribbr's own method so your result sits beside its 78%. Use texts of 1,000 to 1,500 characters and mark each scan 1 if it lands within 15% of the right answer, 0.5 if within 40%, or 0. I haven't scored a set myself.
- Take 3 pieces you wrote before late 2022.
- Generate 3 AI texts on matching topics.
- Run 3 more AI texts through a paraphraser; Scribbr's test caught 50% of mixed and paraphrased texts combined.
- Paste each one into the Scribbr AI detector separately and note the label too.
- Mark every scan 1, 0.5 or 0.
- Add the marks, divide by 9 and compare with 78%.
If one of your own pieces scores as AI, keep your drafts and see how to prove you didn't use AI.
Your essay deserves better than a coin toss
Here's the maddening part. The Scribbr AI detector flagged none of the 60 reworded AI samples in the 2026 Lublin study, yet Lege's audit saw it put an AI percentage on 211 of 303 theses by second-language writers. Misses on one side, nervous scores on the other, and you're stuck in the middle.
That's where EvadeGPT comes to the rescue. Our custom-trained model drives the EvadeGPT AI detector, so you can check your draft on your own terms, then lean on EvadeGPT to rework robotic-sounding lines until they read like you. EVE chat helps with research. Hit the button and try it.
FAQ
Is the Scribbr AI detector accurate?
On human writing, mostly yes. On reworded AI text, no: one 2026 study saw it flag none of 60 paraphrased samples.
Does Scribbr keep the text I check?
Scribbr says "we do not store or share your data." No test has checked that claim.
Which languages does the Scribbr AI checker support?
Scribbr lists English, Spanish, German and French. A 2026 study still got results on Polish: 49.3% accuracy against 38.7% for English.
Can Scribbr check for plagiarism too?
Not with this tool. Scribbr says its AI checker "can only identify AI-generated, AI-refined, and human-written content."
Is a Scribbr score enough to prove AI use?
No. Scribbr admits "no AI model on the market can guarantee 100% accuracy, including ours."


