AI detectors disagree. Provenance shows you where.
Seven independent methods, from watermark tests to a trained classifier, each give their own verdict. When they conflict, you see the conflict instead of one confident-looking number.
Built for research and teaching. It can't detect any vendor's production watermark or classifier, and AI detection isn't solved. Read the limitations.
Less predictable (more human-like)More predictable (more AI-like)
We tried to fix the fence and to paint the shed and to get the goat back in the pen. The goat ate the paintbrush and the fence fell on the shed and the shed fell on the goat. He was fine, and he ate a second paintbrush to prove it.
- Statistical detectorBinoculars score 0.328Likely human-written
- Trained classifier99.95% AI, 90% interval 30–100%Likely AI-generated
- Retrieval ledgerOnly recognizes text this suite has loggedNo match
Modules disagree, abstaining
Pooling B and C would print 64.8% AI. The ensemble abstains instead: when two methods reach opposite verdicts, an average hides the one thing worth knowing.