How to check AI images with Content Credentials

Understand what Content Credentials represent

Content Credentials are provenance information attached to or associated with digital media using the C2PA technical standard. A credential can describe who or what created an asset, which application recorded an action and what edits were declared. The information is placed in a signed manifest so a compatible verifier can detect whether the claims or the asset binding have changed since signing.

The standard creates a chain of assertions rather than a universal truth label. A valid signature can show that a named tool or organization signed particular claims about a particular asset. It does not prove that a depicted event happened, that a caption is honest, or that every earlier step was recorded. Trust still depends on who issued the signing certificate, what the claims say and whether the verifier accepts that issuer.

Credentials may identify AI generation or editing when the producing system records it, but absence is not proof that no AI was used. A camera, editor or publishing platform may not support the standard. Metadata can also be lost when an image is screenshot, copied through an incompatible service or re-encoded. Treat credentials as useful evidence inside a wider verification process.

Inspect the credential and its validation result

Open the original file or the publisher's original link in a verifier that supports Content Credentials. Look first for a validation result. A valid result means the cryptographic checks succeeded for the data being examined; an invalid or unknown result needs investigation. Do not judge only by the presence of a Content Credentials icon or a screenshot of a verification panel, because either can be copied into an unrelated image.

Read the signer or source, creation information and action history. Distinguish a device manufacturer, editing application, news organization and individual creator: each establishes a different kind of context. Review whether the record says an asset was captured, generated, cropped, color-adjusted or composited. The visible history may contain several ingredients or derived versions, and not every ordinary edit indicates deception.

Compare the verified asset with the version you were asked to assess. A manifest can be valid for one file while a social post shows a crop, screenshot or recreation. Check dimensions, visible content, publication URL and timestamps where available. If the tool reports that a manifest cannot be found, record that limited result instead of changing it into a conclusion that the image is authentic or fake.

Use a wider verification workflow

Start with the claim, not only the pixels. Write down what the post says happened, where and when. Locate the earliest available upload and ask for the original file when the decision matters. Search for reporting from the named organization, reverse-search distinctive frames and compare landmarks, weather, shadows, signs and event timing with independent sources.

Then use provenance information to refine the assessment. A trusted capture credential may support a claim about the recording device and subsequent edits. A generative-AI assertion may explain that the asset began in an image model. An edit history may show a harmless crop or a significant composite. Preserve the file and verification output so another reviewer can reproduce what you saw; online copies and trust lists may change.

Separate observations from conclusions. For example: 'the verifier found a valid manifest signed by this application and reports these edits' is an observation. 'The scene is therefore accurate' goes beyond that evidence. For journalism, fraud, legal disputes or safety decisions, use a trained investigator and corroborating records rather than relying on one automated result.

Preserve provenance when publishing your own media

If a supported camera or editing application offers Content Credentials, decide which identity and history information you are comfortable publishing. Review the manifest before export, especially location, creator and organizational details. Use a verified organizational identity only through an approved workflow; a display name by itself should not be treated as strong identity proof.

Keep the signed original in controlled storage and export a distribution copy through a workflow known to retain credentials. Test the actual website, content-management system and social channel because optimization can strip or transform metadata. When the platform supports durable discovery, a verifier may recover associated credentials even when embedded data is removed, but publishers should still confirm what recipients can see.

Document material AI generation and editing in plain language alongside the technical credential. Readers may not have a compatible viewer, and a clear caption gives immediate context. Content Credentials work best when publishers make accurate claims, platforms preserve them and viewers understand their limits; they do not replace editorial accountability.

Quick checklist

• Inspect the original file or publisher link with a compatible verifier.
• Check validation status, signer, actions and asset binding.
• Confirm that the verified file matches the version under review.
• Treat missing credentials as unknown, not proof of authenticity or deception.
• Corroborate the event, place, time and source independently.
• Preserve the file and record observations separately from conclusions.

Primary references

Standards and platform support change. Use these authoritative sources to check the current specification and migration guidance.

• C2PA — Technical Specification 2.3
• Content Credentials — Verify and learn