How to check what is inside an AI image before you share it

Published 21 September 2026 · 8 min read · By Richard Brereton

Before you ask an AI assistant to describe an image, forward it as evidence or publish it publicly, inspect the file you actually have.

The useful question is not simply:

A safer question is:

That change matters because an image can carry several different kinds of evidence—and no single layer gives you a complete history.

Start with the original file

If possible, inspect the original JPEG or PNG rather than a screenshot or a copy downloaded from a messaging platform. Screenshots and re-saves can remove the metadata that was present in an earlier copy.^1

The file you inspect is the file you can make observations about. It may not be the first version of the image, and its metadata may already be incomplete.

Look for Content Credentials

Content Credentials are the user-facing name for a C2PA manifest. They can record information about an asset’s origin, edits, tools and signing information. If an AI tool supports the relevant C2PA actions, its credentials may indicate that the image was generated with AI.^2

But the first result needs careful wording:

• Detected means a credential is present.
• Valid means the credential and asset bindings passed cryptographic checks.
• Trusted means the signer passed the verifier’s configured trust policy.
• Invalid means a signature, hash or other validation check failed.
• Not found means no readable credential was found in this copy.

“Detected” is not the same as “trusted”. A credential can be cryptographically valid while its signer remains outside the trust list used by the verifier. Conversely, a missing credential does not prove that an image is fake or human-made.

The official browser-side C2PA library supports explicit verification contexts and trust-anchor configuration, which is why a careful verifier should expose these states rather than flattening them into a single green tick.^3

Check ordinary metadata as a separate evidence layer

An image can contain ordinary metadata even when it has no Content Credentials. Depending on the device and export path, that may include:

• camera or software information;
• capture or creation dates;
• image dimensions;
• GPS coordinates;
• altitude or accuracy values; and
• device or serial identifiers.

This information can help you understand the file, but it is not automatically trustworthy. Metadata can be copied, changed, stripped or left behind by a tool that does not represent the whole history.

It is also a privacy issue. A photograph taken near your home may reveal more than you intended to share.

Treat location as sensitive by default

Before sharing an image, check whether its metadata includes location information. Exact coordinates may identify a home, workplace, school, regular route or private event.

A privacy-first inspection flow should show an approximate area by default and keep exact coordinates, altitude, capture time and device identifiers behind an explicit local reveal control. If you do not need the exact values, do not include them in a report or certificate that you plan to share.

Understand what missing data means

An image with no Content Credentials is an image with no readable Content Credentials in that particular copy. It is not a verdict.

The Content Authenticity Initiative describes missing credentials as an absence of historical information, not evidence that the image is fraudulent. A screenshot or a photograph of an existing image may also fail to carry the original credential context.^1

That means these statements are too strong:

• “There are no credentials, so it must be fake.”
• “There is no AI marker, so it must be human-made.”
• “The metadata says this, so the whole history is proven.”

Safer wording is:

A practical five-minute checklist

Before sharing an image with an AI assistant or another person:

1. Keep the original file if you have it.
2. Check the file type, size and dimensions.
3. Look for Content Credentials.
4. Read the validation state, not just the presence indicator.
5. Check ordinary metadata separately.
6. Look for GPS, time and device information.
7. Decide whether exact sensitive values need to be revealed.
8. Record uncertainty if credentials are missing or untrusted.
9. Avoid describing the image as definitely real, fake, human-made or AI-made from one weak signal.
10. Share the minimum information needed for the task.

How MUTANT Inspect fits in

MUTANT Inspect is designed around this evidence-first workflow. It checks the selected image locally in the browser and reports:

• MUTANT Signature markers;
• file facts;
• Content Credentials and their validation state;
• readable creator and tool fields;
• ordinary metadata such as location where available; and
• a plain-English interpretation with explicit limits.

The intended experience is local-first: the original image stays in the browser for the inspection flow, sensitive location information is hidden by default and the provenance certificate is generated locally as an unsigned record of observations.

MUTANT Inspect does not claim to decide whether an image is “real”. It helps you understand what this file can show before you rely on it or pass it to another system.

Inspect an image locally with MUTANT Inspect.

Why this matters for AI assistants

An AI assistant usually begins with the pixels: what appears to be in the image, what text is visible and what the scene resembles. That can be useful, but it is different from knowing where the file came from, whether it carries a signed history or whether its metadata exposes a private location.

A provenance check adds a second layer before interpretation. It gives the assistant—or the person using it—more context and a clearer uncertainty boundary.

OpenAI’s current content-provenance guidance makes the same broader point: provenance results are supporting evidence, not absolute proof, and high-stakes decisions should not rely on an automated result alone.^4

The short answer

Check the original file. Look for Content Credentials. Separate detected from valid and trusted. Inspect ordinary metadata for privacy risks. Treat missing information as unknown, not as proof.

The goal is not to make the image tell you everything. The goal is to avoid making the file claim more than it can support.