C2PA vs AI image detection: what each can and cannot prove

Published 21 September 2026 · 6 min read · By Richard Brereton

C2PA and AI image detection are often discussed as if they were competing versions of the same tool.

They are not.

C2PA can record and bind provenance claims when a creator, camera or software tool supports it. An AI detector estimates whether an image resembles material produced by an AI system. Ordinary metadata and watermarks are separate signals with their own limits.

No single layer gives you a complete history of every image.

The four evidence layers

1. C2PA and Content Credentials

C2PA is a standard for recording provenance claims and binding them to an asset through cryptographic mechanisms. Content Credentials may describe origin, edits, tools, ingredients and signing information.^1

A verifier can report whether a credential is detected, whether its cryptographic checks pass and whether its signer is trusted under a configured policy.

This can provide strong positive evidence when the relevant creator or tool recorded a claim and the credential remains attached to the file.

But C2PA is not a universal detector. If a file has no credential, that does not prove it was made by a person. If a file has a valid credential, that does not prove every visual claim is true or that its entire history was recorded.

2. Ordinary metadata

EXIF and other embedded fields can include camera model, software, time, GPS or device information. They can be useful for understanding the file and important for privacy.

Ordinary metadata is not automatically signed. It can be changed, copied or removed. It should be read as context and evidence to examine, not as a complete chain of custody.

3. Watermarks and platform markers

A watermark may indicate that an image passed through a particular tool or platform. Some watermarks are visible; others are designed to be detected by software.

A watermark has a narrower meaning than a complete provenance record. It may be cropped, removed, copied or absent because the exporting tool did not add one. Treat it as one signal rather than a universal authenticity stamp.

4. AI image detection

An AI image detector is usually a classifier or scoring system. It may examine visual patterns, compression behaviour, generation artefacts or other signals and estimate whether the image resembles examples from an AI system.

A detector score is not the same as a signed history. It can produce false positives and false negatives, especially when images are resized, edited, recompressed or generated by unfamiliar systems.

A score should therefore be described as an estimate, not as proof of authorship or origin.

What each layer can support

Why missing C2PA does not mean human-made

A file may have no readable Content Credentials because:

• the original tool did not create them;
• an editor or platform removed them;
• the file is a screenshot or photograph of a screen;
• the credential is in an unsupported format; or
• the image has been re-saved without the original history.

The correct statement is: “No readable credential was found in this copy.” The incorrect statement is: “There is no credential, so the image must be human-made.”^2

Why a detector score does not establish authorship

Even if a classifier gives a high AI-likelihood score, several explanations may remain possible. The image might have been edited after generation, passed through multiple tools or resemble the detector’s reference data for reasons unrelated to its actual origin.

A low score does not prove that a human created the image either. New generation systems, unusual workflows and heavy editing can all affect the result.

This is why MUTANT does not present experimental forensic signals as a confident public verdict. Evidence should be separated by type and explained with its limits.

A practical decision table

How MUTANT Inspect fits

MUTANT Inspect focuses on the evidence layers that can be checked locally in the browser. It reports file facts, Content Credentials and their validation state, readable metadata and privacy-sensitive location information where available.

It does not claim to decide whether an image is definitely real, fake, human-made or AI-made from a weak or missing signal. The product’s useful question is narrower:

See which provenance signals are present in your file with MUTANT Inspect.

The MUTANT Inspect guide explains how to interpret the result labels and why missing information remains unknown.

A safer workflow

Before relying on an image:

1. Preserve the original file if possible.
2. Inspect Content Credentials separately from ordinary metadata.
3. Read the validation state rather than only the presence indicator.
4. Check whether location or device fields need to be removed before sharing.
5. Treat detection scores as estimates.
6. Compare the result with source context and human review.
7. Record what the file does not establish.

This approach is slower than asking one tool for a yes-or-no answer, but it is more honest about the evidence.

The short answer

C2PA records signed provenance claims. Metadata provides technical context. Watermarks indicate a narrower marker. AI detectors estimate resemblance to reference patterns.

They answer different questions. Use them as separate evidence layers, not as interchangeable proof.

Sources

• How Content Credentials work
• Content Authenticity Initiative FAQ
• Official `c2pa-web` browser SDK