The Metadata Gap

The Coalition for Content Provenance and Authenticity (C2PA) specification, now in version 2.3, defines a cryptographic framework for embedding verifiable provenance data into digital media. When present, C2PA content credentials provide an authoritative record of how an image was created — including whether it was generated by an AI system — signed by the producing software or device. OpenAI, Google, Adobe, and Microsoft now embed C2PA metadata in AI-generated outputs. The Nikon Z6 III supports in-camera C2PA signing. The infrastructure for a provenance-first internet is being built.

The problem is that almost nobody uses it yet. As of early 2026, fewer than 5% of images circulating on social media carry C2PA credentials. The specification is complex, tooling is immature, and most platforms strip metadata during upload and re-compression. The provenance-first future is real, but it is not the provenance-first present.

This gap creates an opportunity for detection systems. In the absence of universal provenance, the metadata that does exist — or conspicuously does not — becomes a forensic signal of surprising power.

Consider the EXIF profile of a photograph taken with a modern smartphone. It contains: device manufacturer and model, GPS coordinates, focal length, aperture, shutter speed, ISO sensitivity, white balance parameters, lens identification, timestamps, and often an ICC color profile. This is not merely metadata. It is a forensic fingerprint that is extraordinarily difficult to fabricate convincingly and that AI generators rarely attempt to reproduce.

Conversely, consider the metadata profile of an image generated by Stable Diffusion through a common interface. The PNG output contains tEXt chunks with explicit generation parameters: the prompt, negative prompt, sampling method, CFG scale, seed value, and model hash. The image literally declares its synthetic origin in its file structure. ComfyUI embeds complete workflow graphs. InvokeAI includes its own metadata schema. These are not subtle signals requiring sophisticated analysis. They are confessions embedded in the file header.

Our provenance analysis layer exploits this asymmetry systematically. Before any pixel-level analysis occurs — before any API is called, before any inference cost is incurred — we extract and analyze the complete metadata profile. The analysis examines four signal categories: EXIF structure and content, format-specific metadata (PNG text chunks, JPEG application markers, JUMBF boxes), C2PA content credentials when present, and structural characteristics of the container format itself.

The results inform a triage decision. When metadata is definitive — a PNG with Stable Diffusion parameters, a JPEG with C2PA credentials from DALL-E, or a photograph with rich camera EXIF from a known device — the determination is made immediately, at zero computational cost beyond file header parsing. When metadata is absent or ambiguous, its absence is recorded as a weak signal and the system escalates to more expensive analysis layers.

The structural analysis deserves particular attention. Real camera JPEGs contain a characteristic pattern of APP markers: APP1 for EXIF, APP2 for ICC profiles, often APP13 for IPTC data. AI-generated JPEGs, particularly those saved by Python imaging libraries or web applications, typically contain only APP0 (JFIF) and sometimes APP1. The marker count and composition are themselves a forensic feature — a signature of the software stack that produced the file, distinct from the image content.

In our testing, provenance analysis alone resolves approximately 15–20% of cases with high confidence. These are the easy cases — images that announce their origin in their file structure. But even in cases where metadata is inconclusive, the information it provides (or fails to provide) shifts the prior probability in a direction that meaningfully influences downstream analysis. A JPEG stripped of all metadata is not the same as a JPEG with full camera EXIF. Both may require further analysis, but the forensic context they provide is different, and that context matters.

We view provenance analysis not as a detection method but as a triage function — the zero-cost first pass that prevents unnecessary expenditure on cases where the answer is already encoded in the file. As C2PA adoption increases and more generators embed verifiable credentials, this layer will become more powerful, not less. The future of detection may well be a world where the hardest cases are the only cases that remain, because the easy ones are solved by the images themselves.