AI images are generated, not stamped from one template
Modern diffusion systems usually begin with noise in a latent representation, then repeatedly denoise it under the influence of a prompt, model and generation settings. The process creates statistical regularities in pixels, but those regularities change across model families, resolutions, edits and exports. The foundational latent diffusion paper describes this generation process.
Source: Latent Diffusion Models, CVPR 2022
Four evidence layers answer different questions
Visual patterns
A versioned classifier looks for learned spatial and frequency correlations. It provides a probabilistic clue, not a signature.
Generator metadata
Software names, prompts, seeds, models and workflow graphs can identify a generation path, but ordinary metadata is editable.
Signed provenance
A C2PA manifest binds assertions to file bytes and records validation status, signer information, actions and ingredients.
Provider watermark
A proprietary signal such as SynthID needs the provider’s compatible verifier. A generic metadata scan cannot substitute for it.
Why visual classifiers can disagree
Research detectors learn correlations from known training examples. New generators, heavy compression, crops, illustrations and unusual image processing can shift those correlations. Spectral-detection research documents useful artifacts and the difficulty of generalising across generators. A visual score must therefore name its model, version and thresholds.
Source: Any-Resolution AI-Generated Image Detection, CVPR 2025
Why metadata is useful but not immutable
PNG text chunks, EXIF, XMP and application-specific records may contain an explicit generator, prompt or ComfyUI graph. That is strong file-specific evidence when present. It can still be removed by a screenshot, re-encoding, a social platform or an editor configured to strip metadata.
How C2PA Content Credentials are built
A C2PA manifest combines assertions, a claim, a cryptographic signature and a hard binding to the asset. Validation checks structure, assertion hashes, the file binding, signature, time stamp and certificate policy. The current C2PA 2.4 specification defines provenance and tamper evidence; it does not declare whether the depicted event is true.
Source: C2PA Technical Specification 2.4
| C2PA state | What it establishes | What it does not establish |
|---|---|---|
| Trusted / valid | The supported manifest and file binding validated under the selected trust policy. | That every assertion is objectively true. |
| Valid but untrusted | The structure and binding validate, but the signer is outside the selected trust list. | That the signer is malicious or that the image is fake. |
| Invalid / tampered | A signature, assertion or file-binding check failed. | Which edit or person caused the failure. |
| Absent | No supported manifest was found in this copy. | That the image came from a camera or a human. |
SynthID and other invisible watermarks
Google DeepMind describes SynthID as an imperceptible signal embedded in generated content and read by a compatible detector. Google also documents limits: a negative SynthID result applies to supported Google AI systems, not every generator. This site labels SynthID as not tested unless a compatible provider check is actually available.
Source: Google DeepMind — SynthID
Why our layered report is more useful than one percentage
- Identify the container before interpreting fields.
- Read supported PNG, EXIF, XMP and workflow records without treating filenames as proof.
- Validate C2PA separately with a conforming reader and preserve validation states.
- Run the pinned visual model as a separate estimate with published thresholds.
- Display contradictions instead of averaging unrelated signals into a confident-looking number.
How to read the result
| Result | Narrow interpretation |
|---|---|
| AI evidence found | An explicit supported generator, workflow or AI source declaration was found. |
| AI-like visual signal | The visual model crossed its conservative display threshold; false results remain possible. |
| Inspectable signals | Metadata or provenance exists, but it does not by itself declare AI generation. |
| Provenance invalid | A supported C2PA validation check failed and should be reviewed. |
| Inconclusive | No supported decisive signal was found in this copy. |
Which tools can verify the layers?
Use a C2PA validator or Content Credentials inspector for signed provenance, ExifTool for broad metadata inspection, ComfyUI for embedded ComfyUI graphs, AUTOMATIC1111 PNG Info for its saved parameters, and the provider’s own verifier for proprietary watermarks. Tool compatibility depends on format and export path.
Frequently asked questions
Can AI metadata be removed?
Yes. Ordinary metadata and workflow records can be stripped. C2PA tampering is detectable when the manifest remains available, but the whole manifest can also be omitted from a later copy.
Does missing C2PA mean a real photo?
No. It means this copy has no supported C2PA manifest.
Can an AI tool edit a real photograph?
Yes. Provenance can describe ingredients and actions, while a visual model may react to only part of a composite.