How accurate is the ZeroTell AI Detector?
On our held-out test set of 1,826 documents in 10 languages, the detector flagged 4 of 671 human-written documents as AI (0.6%) and recognized 90.6% of the AI-written ones. On text that another model paraphrased it does much worse, and no score is proof of who wrote something.
The short answer
ZeroTell's AI detector sends every sentence of a text to Google's Gemini 3.8 Flash model with our own scoring instructions. A sentence that scores 50% or higher is highlighted as AI-like, and the document's score is the share of its text that sits in highlighted sentences. On October 28, 2026 we measured the exact model and settings the site uses on three labeled test sets. Each number below comes straight from the evaluation report our test script produced; we only quote groups with at least 200 documents or sentences.
Two results matter most. Human writing is rarely flagged: 0.6% of human documents in the held-out set and 0.9% of the essays written by non-native English students. But AI text that was paraphrased by another model is missed far more often in our own data: only 71.1% of such documents were caught in the held-out set. On the public MAGE benchmark the same kind of text was caught far more often (96.9%), so results depend strongly on the text.
Results by test set (document level)
A document counts as AI-written when its score is 50% or higher, that is, when at least half of its text is in highlighted sentences. A mixed test document counts as AI-written when at least half of its text was written by AI. False positive rate = human documents wrongly flagged as AI. Detection rate = AI documents correctly flagged. AUROC (0.5 = guessing, 1.0 = perfect separation) does not depend on the threshold.
| Test set | Documents | Human documents flagged as AI | AI documents detected | AUROC |
|---|---|---|---|---|
| Held-out test set (10 languages) | 1,826 | 0.6% (4 of 671) | 90.6% | 0.954 |
| Student essays (non-native English writers) | 555 | 0.9% (2 of 222) | 98.2% | 0.992 |
| MAGE public benchmark (out of distribution) | 1,599 | 0.3% (1 of 399) | 96.6% | 0.990 |
What kind of AI text gets missed
Detection rate at the 50% threshold, split by how the AI text was produced. Groups with fewer than 200 documents are left out.
| Test set | AI text, unedited | AI text paraphrased by another model |
|---|---|---|
| Held-out test set | 97.3% | 71.1% |
| MAGE public benchmark | 96% | 96.9% |
Sentence-level results by language
The detector highlights a sentence when its score is 50% or higher. Held-out test set; every language below has at least 200 sentences.
| Language | Sentences | Human sentences highlighted | AI sentences highlighted | AUROC |
|---|---|---|---|---|
| English | 8,485 | 0.6% | 91.1% | 0.986 |
| Chinese | 5,762 | 0.1% | 92.3% | 0.980 |
| Japanese | 1,588 | 1.6% | 97.7% | 0.996 |
| Korean | 877 | 0% | 87.4% | 0.987 |
| Spanish | 1,166 | 0% | 96.2% | 0.984 |
| German | 1,109 | 1.5% | 73.5% | 0.901 |
| Italian | 1,038 | 0% | 99.1% | 1.000 |
| Portuguese | 980 | 0% | 91.4% | 0.993 |
| Malay | 1,205 | 0% | 81.6% | 0.982 |
| Turkish | 1,413 | 1.6% | 88.6% | 0.986 |
The three test sets
Held-out test set (1,827 documents, 23,635 sentences): a fixed share of our own labeled data that we never use to develop the detector. Documents are assigned by source group, so no source text from our development data, and none of its AI rewrites, appears here. Human texts come from published sources, most of them written before AI writing tools were common (news, encyclopedia and web articles, human answers to questions); AI texts include unedited model output, text paraphrased by a second model, humanized drafts and mixed documents, in 10 languages. Most AI documents in this set were written by Gemini and GPT models.
Student essays (555 documents): essays by non-native English-speaking students from the ELLIPSE corpus, paired with AI essays written on the same prompts and AI essays that were then paraphrased or humanized. This set is used only for evaluation, never to develop the detector.
MAGE benchmark (1,600 documents): a public research benchmark for machine-generated text that we did not use to develop the detector. It includes AI text from model families that are not in our own data and a large share of paraphrased AI text. Because the benchmark is public, the underlying Gemini model may have seen parts of it during its own training, so read its results as an optimistic case.
Limits you should know about
These are results on our test sets, not a guarantee for any particular text. Short texts, heavily edited drafts, paraphrased AI text and writing from models we did not test can all score differently.
A low false positive rate is not zero. If you are flagged, treat the score as a reason to look closer, not as proof — see our Academic Integrity page for what to do.
The numbers are tied to one model and one version of how text is sent to it. When we change either, we run the evaluation again and replace this report.
How we measured
An evaluation script sends every test document through the same detector the site uses, then compares each document's and each sentence's score with its known label. It reports results at two thresholds (50%, the one the site uses, and 70% for comparison), per language, genre and source, and marks any group with fewer than 200 items as too small to quote. This page shows only those larger groups. Documents that Google's content filter refused to process have no score and are left out of the numbers: 1 in the held-out set, 0 among the student essays and 1 in MAGE. Every other document was scored.
Detector accuracy FAQs
On our held-out test set of 1,826 documents, it flagged 0.6% of human documents as AI and detected 90.6% of AI documents (AUROC 0.954). Results vary by test set, and paraphrased AI text is caught far less often.
Rarely on our test sets: 0.6% of human documents in the held-out set and 0.9% of essays by non-native English students. That is not zero, so a score is never proof on its own.
Less reliably. In our held-out set, AI text paraphrased by another model was detected 71.1% of the time, against 97.3% for unedited AI text. On the public MAGE benchmark the paraphrased rate was 96.9%, so it varies a lot with the text.
No. We ran it ourselves on our own test sets and a public benchmark. We publish the method, the sets and the limits so you can judge the numbers, and we rerun it whenever the model changes.
No. Detection is probabilistic. Use the score as one signal alongside drafts, version history and a conversation with the writer, as our Academic Integrity page explains.
Check a text with the ZeroTell AI Detector
See which sentences look AI-written and why — then judge for yourself.
Open the AI Detector