EU AI Act Annex IV style · FTC-safe disclosure
Product technical file & known limitations
Detection pipeline
- Input ingestion (file, URL, text; optional SHA-256 hash)
- Modality preprocessing & detector ensemble
- C2PA / watermark / registry cross-check (Verify Suite)
- Signed PDF + certificate key + JSON-LD compliance label
Known limitations
- Probabilistic scores — false positives and false negatives occur
- Not legal proof, government approval, or court-expert certification
- Novel generators, compression, and adversarial edits reduce accuracy
- Annex III high-risk sole automated decisions prohibited (AUP)
Accuracy & benchmark transparency
Internal hold-out evaluation at default operating threshold. Probabilistic signals only — not legal proof. Independent third-party benchmark in progress. Published 2026-06-16 · internal holdout · third-party validation pending
| Modality | n | Precision | Recall | F1 | FPR | FNR | Notes |
|---|---|---|---|---|---|---|---|
| image | 248 | 86% | 81% | 83% | 14% | 19% | Strong on common diffusion outputs; degrades on heavy JPEG recompression and aggressive crops. |
| text | 312 | 91% | 76% | 83% | 9% | 24% | Longer samples more stable; human+AI hybrid edits increase false negatives. |
| audio | 186 | 83% | 77% | 80% | 17% | 23% | Voice clones and TTS; telephony noise reduces recall. |
| video | 142 | 80% | 74% | 77% | 20% | 26% | Frame sampling pipeline; clips under 3s less reliable. |
Download machine-readable benchmark JSON · Full methodology
