Possible synthetic indicators
Paired with model/version, raw score, limitations and review context.
Trust center
Show the limits.
Earn the decision.
Claims, lane maturity and data handling stay visible before integration.
See the measured accuracy Inspect the contractThe bounded evidence, first.
The current lane is aidetect-v16-image-384-v1+owenslab+siglip-probe-routed-v11+dualview-orig-v1. Measured through that lane: 78.1% of 653 AI images drew a strong warning, and 4.0% of 795 real ones drew a false one. Video is judged on 15 frames: 82.6% caught on 23 videos. These are bounded slices, not overall accuracy or proof of origin. Read the samples and limitations.
Lane maturity
No universal detector. Different media, models and failure slices require different policies.
| Lane | Status | Public use | Key limitation |
|---|---|---|---|
| Image | Core warning lane | Human-review support | Art, screenshots and unseen generators can fail |
| Sampled video | Core aggregate | Frame-level triage | Not temporal deepfake analysis |
Proof of origin
Do not say definitely AI, definitely human, authentic, guaranteed accurate or calibrated probability without independent evidence.
Provenance is another facet, not a verdict.
Content Credentials can be present, absent or fail validation. Their absence is not evidence that media is synthetic.
- Source declaration
- Credential validation state
- Duplicate and hash history
- Detector context
- Human decision
Draft retention defaults
These terms require legal and operational review before production publication.
- Raw media
- 24 hours in active systems
- Review dossiers
- 30 days
- Backup deletion lag
- Up to 14 days
- Training
- No customer-media training by default