Start with provenance and metadata
If a valid C2PA credential already names a trusted AI vendor or product, that can be one of the strongest signals available. If provenance is absent, review EXIF and XMP metadata for capture history, editing traces, and unusual gaps.
Look for visual inconsistencies
Synthetic textures, inconsistent reflections, deformed text, implausible shadows, and broken object geometry can all matter. But visual clues alone are not enough, because heavy editing and compression can create similar artifacts.
Use detector scores as risk signals
Commercial detectors and visual reasoning systems are best used to prioritize review, not to declare absolute truth. If multiple layers point in the same direction, confidence increases. If they disagree, that is usually a sign to escalate.
Run the workflow on a real image.
If you want to see these signals together instead of reading them in isolation, start with a free account and review one report end to end.