For writers taught to write formally
AI detectors give one verdict. This shows you where they disagree, before anyone accuses you.
AI detectors can misread careful, formal writing as machine-written. In our tests, a commercial detector flagged 13% of real student essays written before ChatGPT existed, and 60% of human-written scientific abstracts. A single score can't tell you which kind of mistake you're looking at.
This tool doesn't hand you one more score to argue about. Paste in a piece of text and it runs through several independent detectors — a fine-tuned classifier, statistical detectors, and a commercial detector (Sapling) — at once, and shows you exactly where they agree and where they don't.
Detector results are probabilistic signals, not proof of AI use or human authorship. This tool is not a substitute for an institution's own academic-integrity process.
Same paragraph, different detectors, different answers
- RoBERTa classifierHuman
- Log-perplexity detectorAI-generated
- GPTZeroUnavailable
Illustrative example — not a live result.
Run the scan
Paste in a paragraph or two. It runs through every detector we have wired up and shows you exactly where they land.
What we found
Preliminary results, November 2026. Not peer-reviewed.
- A commercial detector (Sapling) flagged 13% of 170 genuine student essays written before ChatGPT existed. Our own fine-tuned classifier flagged 18%.
- In a pre-registered test, Sapling flagged 36 of 60 human-written scientific abstracts from before 2020 (60%) as AI-generated, compared with 12% of student essays. Text length didn't explain it: student essays cut to abstract length were flagged 8% of the time. The other three detectors flagged almost none of the abstracts.
- Detection depended heavily on which AI wrote the text. Two free statistical detectors caught 70-93% of GPT-written text but 5% or less of Gemini-written text.
- We expected non-native English writers to be flagged more often. Across two studies the difference leaned that way but was not statistically reliable at our sample size.
Human essays: Write & Improve corpus and LOCNESS (Centre for English Corpus Linguistics, UCLouvain), via BEA-2019. Abstracts: arXiv (public domain).