Content credentials and C2PA: labels that show who made what

As AI-made images, audio and video become harder to tell apart from human work, more cameras, editing apps and AI tools now attach a provenance record to the files they produce. This page, part of our guide to AI and copyright, explains what those records are, what they can and cannot show, how they differ from watermarks, and what they mean for copyright.

The standard behind them is published by the Coalition for Content Provenance and Authenticity (C2PA). "Content Credentials" is the plain-language name for its signed records.

01How a content credential works

The C2PA and Content Credentials Explainer for specification version 2.2 describes the process in steps. In plain terms:

1. Creation. A camera, app or AI tool that supports the standard records facts about the file as it is made.
2. The manifest. Those facts are stored as "assertions" in a C2PA Manifest, the technical name for a content credential. Assertions can cover origin, edits and use of AI.
3. Signing. The software or device signs the manifest with a digital certificate, and a cryptographic hash ties it to the exact content.
4. Embedding. The credential usually travels inside the file. It can also be stored separately and found again through a watermark or fingerprint, which the standard calls a soft binding.
5. Checking. A credential-aware viewer checks the signature and hash and shows the history, often behind a small icon.

When a file is built from other files, each source is recorded as an "ingredient", creating what the explainer compares to a family tree of provenance. When an AI system performed an action, the explainer says it is identified through a field called digitalSourceType.

The C2PA site listed version 2.5 as the newest specification when this page was reviewed in October 2026. The core ideas above appear across recent versions.

02What it proves and what it does not

The C2PA explainer is unusually frank about limits. It says content credentials "do not provide value judgments about whether a given set of provenance data is 'true'". What they show is narrower:

Two more limits matter in practice. First, provenance is not always complete: if a file is cropped in a tool that ignores the standard, that step may go unrecorded. Second, the explainer confirms that provenance metadata can be removed. Many websites and apps also strip metadata on upload.

The explainer also warns against treating missing credentials as suspicious, because adding provenance is optional. A content credential is evidence about a file's history, not a verdict on its truth or its ownership.

03Watermarks versus credentials

The words get mixed up, so here is the difference.

• Visible watermark: a logo or text over the image. Easy to see, easy to crop out, and carries no verifiable history.
• Invisible watermark: a pattern hidden in the pixels or sound. It can survive some copying and editing, and can point back to a record. On its own it holds little information.
• Content credential: a signed, structured record of origin and edits. It holds much more information but can be stripped from the file.

The C2PA design combines them. The explainer calls this a "durable" credential: a hard binding (the cryptographic hash) plus a soft binding (a watermark or fingerprint), so a stripped credential can be found again in a cloud store.

The law is starting to expect both kinds of marking. Article 50(2) of the EU AI Act (Regulation 2024/1689), applicable from August 2, 2026, requires providers of AI systems that generate synthetic audio, images, video or text to mark outputs "in a machine-readable format and detectable as artificially generated or manipulated." The Act does not name a single technology. For a policy view of how C2PA fits that duty, see this analysis of watermarking and C2PA as a compliance floor.

04Credentials and copyright

A content credential is not a copyright registration and does not create or transfer any rights. Copyright exists in an original work from the moment it is fixed, whether or not it carries a credential.

Credentials can still help in three ways:

• Showing human contribution. The US Copyright Office's Part 2 report of January 29, 2025 concluded that copyright covers human expression "perceptible in AI-generated outputs" and creative modifications of them, but not purely AI-generated material. A credential that records your edits can support your account of what you did. For other kinds of evidence, see how to prove human authorship.
• Disputes and takedowns. If your work is copied, or wrongly called AI-made, a credential can add weight to your side. Our page on AI-related takedowns explains those processes.
• Training preferences. The explainer notes that credentials can include assertions about "whether they wish to allow its use for AI training". That is a signal of your wishes, not a lock on the file.

One more caution: the explainer says the core specification does not directly address the identity of individual creators, which is handled by extensions. So a credential may tell you a file came from a particular app without naming the person who made it.

05Adding credentials to your own work

1. Check your tools. Some cameras, phones and photo or video editors can attach content credentials. Look in export or settings menus for "content credentials" or "C2PA".
2. Turn on AI disclosure in the credential where your tool offers it, so the record matches what you say publicly. Our guide to writing an AI disclosure statement covers the public side.
3. Add identity only if you want to. Linking your name or accounts is optional and has privacy trade-offs the C2PA explainer itself mentions.
4. Keep originals. Platforms may strip credentials on upload. Your own signed originals remain your best record.
5. Verify before you rely on it. Use a credential-aware viewer to check that the record says what you expect.