Methodology
ImageVerity is designed to combine provenance, metadata, detector scores, and visual reasoning into one review workflow. The product is meant to help teams make better operational decisions, not to pretend one score can settle the truth by itself.
ExifTool reads EXIF, XMP, IPTC, ICC, maker notes, and other metadata fields that may surface capture details, editing traces, tool-chain hints, or unusual metadata gaps. This helps reviewers understand file history, but metadata can be missing, stripped, or edited, so it is never treated as final proof.
c2patool checks whether a C2PA Content Credential exists, who signed it, and whether the recorded provenance chain still verifies. C2PA can provide strong evidence about origin and history, but provenance alone cannot prove that the depicted content is true, accurate, or factual.
Sightengine contributes a pixel-based AI-generation confidence score. Because this layer does not rely on EXIF, C2PA, or watermarks, it can still provide a useful signal when metadata has been stripped, but the result remains probabilistic.
Hive can be used as an additional review layer for AI-generated image/video and deepfake risk, including model classifications and confidence-style outputs. ImageVerity treats it as a second opinion for prioritization and escalation, not as an automatic final decision.
VRE adds multi-modal visual reasoning around suspicious textures, lighting, geometry, reflections, hands, text artifacts, object boundaries, and scene consistency. It is used to make the review path legible, not to claim that visual reasoning alone can settle the truth.
Interpretation rules
No single layer is treated as complete proof.
Conflicting evidence is a reason to escalate, not to overclaim.
Trusted provenance matters, but its absence is not a verdict by itself.
The system is built for operational review, moderation, and risk judgment.
The system is not presented as a final legal or factual adjudicator.
Best-fit workflows
The methodology is strongest in newsroom verification, e-commerce risk control, trust and safety review, and any workflow where a team needs a readable evidence chain before escalating a suspicious image. It is especially useful when a human reviewer needs context, not just a raw score.
Evidence boundaries
Sightengine and Hive signals help prioritize review and identify images that deserve closer attention. They should not be described as 100% proof that an image is AI-generated.
A valid C2PA chain can be a strong provenance signal, but missing credentials do not automatically mean the image is fake. Some platforms and editing tools may remove or fail to preserve provenance.
EXIF/XMP fields can reveal useful capture and editing context, but metadata can be absent or manipulated. It is strongest when it agrees with provenance, detector, and visual evidence.
The practical position is simple: ImageVerity turns several technical signals into a readable review workflow, while keeping the final interpretation cautious.
Official references
Start with a free account, run a detection, and inspect how the signal layers combine before deciding what belongs in manual review.
