For publishers, editors & content agencies
Review a draft
before it goes live.
Check submitted writing for signs of AI with tropa-3. Find passages to look at more closely, discuss them with the writer and keep a report of your findings.
Try text detection free. Sentence-level analysis is available with Pro.
From submission to a documented review.
From a passage to a closer look
Workflow illustrationThe study compares interviews from two communities. The findings suggest that local context shapes participation. The next section examines the evidence behind this conclusion.
Read the overall result
Start with the document-level AI assessment.
Inspect individual sentences
Use sentence-level scores to choose passages for a closer look.
Make an informed review
Check sources, drafting history and context before drawing a conclusion.
Start with the submitted draft
Paste a continuous passage, or sign in to upload a PDF or DOCX. Check the original wording in its original language.
Read the passages behind the score
On Pro, run sentence-level analysis on pasted text. Review flagged passages alongside the brief, sources and the writer’s explanation.
Keep the result with your notes
Export a PDF report. Use the findings to guide follow-up questions and document your editorial decision.
AI detection is a signal for review. It does not prove authorship, verify sources or replace your editorial judgment. Both false positives and missed AI text are possible.
Try it with your own writing.
Use a continuous passage of at least 100 words where possible. A known human draft and a known AI draft can help you evaluate how the tool behaves on your material.
Inside the text detector
A score is the starting point.
Our own text model, tropa-3, gives you signals to investigate and results you can document. We publish how we test it, including the cases it gets wrong.
Meet tropa-3Find the passages to review
Sentence-level scores help you locate the signal in pasted text. Available with Unlimited.
Look for signs of editing
tropa-3 can flag possibly humanized text and report hidden characters that may affect a check.
Keep a record of the result
Export a PDF report to share your findings or keep them with your review notes.
Our published benchmark · September 2026
The same texts. Results you can inspect.
513 texts scored by tropa-3, GPTZero, Pangram and ZeroGPT. Run by us, with AI text from the same models our detector was trained on. Results apply to this test set.
Compare the published results
Both charts use a fixed 0–100% scale. Confident AI labels only.
Raw AI texts detected
159 texts · Higher is better
- tropa-395.6%
- GPTZero61.6%
- Pangram44.0%
- ZeroGPT21.4%
Human texts wrongly flagged
247 texts · Lower is better
- tropa-30% · 0/247
- GPTZero0% · 0/247
- Pangram0% · 0/247
- ZeroGPT0.8% · 2/247
tropa-3, GPTZero and Pangram flagged none of the human texts in this sample. GPTZero and Pangram gave fewer raw AI texts a confident AI label. These results describe this test set, not general accuracy.
| Detector | AI texts detectedof 159 raw texts | Human texts flaggedof 247 human texts |
|---|---|---|
| tropa-3Our model | 95.6% | 0 / 247 |
| GPTZero | 61.6% | 0 / 247 |
| Pangram | 44.0% | 0 / 247 |
| ZeroGPT | 21.4% | 2 / 247 |
Tested 24–25 September 2026. Counts each tool's confident AI label; mixed or partly-AI labels are excluded. Including the mixed label raises GPTZero here to 84.3% and Pangram to 81.1%. The AI texts come from models tropa-3 was trained on, which may favor it. This table excludes paraphrased and humanized text.
Inspect every text and verdict on Hugging Face ↗Make it part of your review process.
For your own reviews
Pro includes unlimited website checks, sentence-level text analysis and PDF reports for $14.99/month. Standard checks and reports are also available using one-time credit packs.
See plans and credits →For a recurring workload
Tell Patrick about your publication, languages and submission volume. Discuss a pilot with a sample of your material, or explore the API for your existing tools.
Discuss an editorial pilot →