What Is Media Provenance?

What is media provenance, in plain terms?

Media provenance is the documented chain of custody for a piece of digital content. For a photograph it might record the camera that captured it, the time of capture, the editing application that cropped it, and the newsroom that published it. For an AI-generated image it might record the model or service that produced it. The word comes from the art world, where provenance is the ownership history that helps establish whether a painting is what it claims to be.

The Coalition for Content Provenance and Authenticity (C2PA) gives the technical definition most of the industry now uses: provenance is the history of an asset, as represented by its C2PA manifests and the manifests of its ingredients. An ingredient is any earlier asset that went into the current one, such as two photos combined into a composite. Each ingredient carries its own history, so a finished file can hold a provenance tree rather than a single line.

NIST's November 2024 report on synthetic content, NIST AI 100-4, groups provenance under "provenance data tracking": recording information about content's origins and history so that others can later judge its authenticity, integrity and credibility. That framing matters because provenance is something added at creation or edit time. A file that was never given a record cannot be checked against one.

How does signed provenance work under C2PA?

A C2PA manifest, which the public-facing name calls a Content Credential, is a bundle of assertions about the asset: actions taken, tools used, ingredients, and optionally an AI-use declaration. The software or hardware that performed the operation signs the bundle with its private key. A hard binding, a cryptographic hash of the content, ties the manifest to the exact bytes it describes, so any later change to the pixels or the record breaks validation.

Validation checks three narrow things: that the manifest is well formed, that the hard binding is intact, and that the signing certificate chains to a trust anchor the validator accepts. The C2PA Trust List, run under a conformance program launched on 4 June 2025, is one such anchor. C2PA's July 2026 deployment guidance reports that 54 products from 31 companies had passed conformance by 14 May 2026. A viewer such as verify.contentauthenticity.org reports a file as valid (intact, signer not necessarily trusted), trusted (intact, signer on the Trust List) or carrying no credentials.

Signed provenance can also make ordinary metadata tamper-evident. The C2PA explainer describes a mapping from EXIF, XMP and IPTC fields into assertions, which means a capture time or a digital source type such as IPTC's trainedAlgorithmicMedia ("Created using Generative AI") can be covered by the signature instead of sitting loose in the file. Our page on metadata and media provenance explains that difference in detail.

What does provenance prove, and what does it leave open?

A valid, trusted credential proves that a known signer made specific statements about this file and that neither the file nor the statements changed afterwards. That is useful evidence. It answers who vouches for the history and whether the history was altered.

It does not answer whether the depicted event happened. The C2PA explainer states plainly that Content Credentials make no value judgment about whether provenance data is true, and that provenance information alone cannot establish whether content is accurate or factual. A real camera can photograph a staged scene. A legitimate editing tool can sign a misleading composite. A signer can make a mistake or have its key compromised. Provenance narrows the question to whether you trust the signer and whether its assertions cover the claim you care about.

Provenance can also be incomplete without anyone acting in bad faith. The explainer notes that history breaks when an asset passes through a tool that does not support Content Credentials. NIST AI 100-4 adds that metadata is often stripped when files are shared on social platforms, for privacy or size reasons. Reading what an empty result means is a separate skill, covered in what missing provenance means.

How is media provenance different from deepfake detection?

Provenance and detection look at different evidence. Provenance reads a record someone attached to the file. Detection examines the content itself, the pixels, frames or audio samples, for traces of generation or manipulation. Provenance needs cooperation at creation time; detection works on any file, including one that a hostile actor deliberately stripped. Detection returns a probability and needs a threshold; a validated credential returns a defined status.

In practice the two answer each other's blind spots. A credential that says "captured on a phone, no edits" is strong context, and a detector finding face manipulation on the same file is a reason to examine the signer. A file with no credential at all, which describes most media in circulation today, can only be assessed from its content and context. We compare the two approaches side by side in C2PA provenance and forensic detection.

How should an investigator check a file's provenance?

Work on the original file whenever possible. C2PA's deployment guidance warns that screenshotting or re-exporting through tools that remove metadata can strip embedded credentials, so a copy pulled from a chat app may tell you nothing even if the source was signed.

• Run the file through a C2PA validator and record the exact status: trusted, valid but untrusted, invalid, or no credential found.
• Read who signed it, when, with which tool, and whether an AI-generation or AI-edit assertion is present.
• Open the ingredient tree and check whether the asset under review is the one the claim is about, not an earlier or later version.
• Test the real-world claim (place, date, people) against evidence outside the credential, such as first-party records or other footage.
• If there is no credential, move on to source tracing, reverse image search and forensic analysis rather than treating absence as a verdict.
• Log the validator name and version, the trust list used and the result, so another reviewer can repeat the check.

Where does NeuroDetect fit when provenance is missing or disputed?

NeuroDetect is Neural Defend's forensic media workspace for exactly the cases provenance cannot settle. An analyst uploads an authorized image, audio file, video or document into a case. The analysis keeps model and version context, and detailed views (the original, an explainability heatmap and attention views) point to the regions, frames and audio segments that drove the result. An LLM-assisted summary explains the findings in plain language and links each statement to evidence, while the analyst makes the final call.

The output is built to sit next to a provenance check. NeuroDetect's downloadable forensic report records observations, uncertainty, provenance, model context and analyst notes, so a reviewer can show both what the file's history claimed and what its content showed. The detection models come from ARCAS 1B, Neural Defend's proprietary deepfake detection model; the ARCAS paper reports pooled accuracy of 99.99% on GenImage and 97.65% on OpenFake. Those are benchmark figures on research datasets, so treat them as a reason to test NeuroDetect on your own case material rather than as a guarantee for any single file.

Teams that receive media at volume, such as a claims desk or a platform trust team, can run the same analysis through NeuroVerify's API and send only ambiguous or disputed files to an analyst in NeuroDetect.