We show you what other detectors hide.
This is a demo for a truly, 100% interpretable and faithful AI text detector. It will show you, and we hope it will help you understand, the exact reasons for why any decision was made; no gatekeeping, all fully interpretable, we promise!
Test AUROC
0.9465
95% CI 0.941–0.952
Live concepts
379
of 2048; 1669 never fire
Embedder
meta-llama/Llama-3.2-3B
frozen, per sentence
Train documents
50,000
~2.9M sentences
Warning: this detector is experimental
Please don't use this detector to make any real-world decisions. We're working on it: results may change without notice, and current results haven't been properly validated. Take this as a demo of what an interpretable detector could be, not a final, validated implementation.