The Asymmetry of Metadata Evidence
When investigating whether a digital image was generated by an AI model, metadata analysis is often the first technique employed. However, metadata has distinct evidentiary characteristics that must be understood carefully.
Scenario A: Positive Metadata Signals (High Confidence)
When an image contains:
- A valid, untampered C2PA manifest signed by a known provider (e.g., OpenAI, Adobe) stating generative creation; or
- Embedded PNG generation chunks containing Stable Diffusion prompt strings, samplers, and seeds;
We can conclude with high confidence that the file originated from an AI pipeline, provided the cryptographic signature or proprietary chunk format is intact.
Scenario B: The Absence of Metadata (Zero Guarantee)
Crucially, the absence of AI metadata does NOT prove an image is a genuine photograph.
Why metadata disappears naturally:
- Social Media Stripping: Platforms like Twitter/X, Instagram, WhatsApp, and Facebook routinely strip all EXIF, XMP, and PNG chunks to save bandwidth and protect privacy.
- Screenshots: Taking a screenshot of an AI-generated image produces a new file with clean default OS metadata.
- Format Conversion: Converting a PNG to JPEG or WebP often discards non-standard chunks.
The Golden Rule of Digital Forensics
Presence of metadata can be diagnostic; absence of metadata is inconclusive.
Reliable forensic pipelines combine metadata inspection with visual artifact analysis and provenance verification rather than relying on a single signal.