A photo arrives in your inbox. It could be a camera capture, a Photoshop composite, or the output of an image generator, and the pixels alone will not tell you which. C2PA is the standard that lets the file answer the question itself, by carrying a signed record of where it came from and what was done to it. This guide explains what that record contains, who writes it today, how to read one, and what it can and cannot prove.
What C2PA is
C2PA stands for the Coalition for Content Provenance and Authenticity. It is an open technical specification for attaching provenance information to media files: images, video, audio and documents. The coalition was founded in 2021 by Adobe, Arm, the BBC, Intel, Microsoft and Truepic. Google, OpenAI and Meta joined the steering committee in 2024, where Sony also sits, and Leica, Canon and Nikon have shipped the standard in cameras. Anyone can read the specification and implement it; the reference software is open source.
The consumer-facing name for a C2PA record is Content Credentials, and the small "CR" pin you may have seen on images is its icon. The two terms describe the same thing from different sides: C2PA is the standard, Content Credentials is what a file carries when a tool has applied it.
The idea is older than generative AI. Newsrooms have needed a way to trace a photo back to its source for as long as photos have been digital, and metadata fields like EXIF and IPTC were the first attempt. Those fields can be edited by anyone with a text editor. C2PA replaces them with a record that is cryptographically signed, so a change to the file or to the record shows up as a broken signature.
What a manifest contains
The record inside a file is called a manifest. A manifest has three parts.
Assertions are the statements the signer makes about the content. A camera writes a capture assertion: which model, at what time, with what lens. An editor writes action assertions: cropped, colour adjusted, a layer added. An AI generator writes a creation assertion that names the model. The most important assertion is the content hash, a fingerprint of the image bytes at the moment of signing. The specification calls this the hard binding, because it ties the record to one exact version of the pixels.
The claim bundles the assertions and names the tool that generated them.
The signature seals the claim with a private key. The matching certificate travels inside the manifest, so a validator can check the signature offline and see who issued the certificate. Any change to the image or to the assertions after signing makes the hash or the signature fail.
When an image is edited in C2PA-aware software, the earlier manifest is not overwritten. It is kept as an ingredient of the new one, so a file can carry a chain: capture, then edit, then export, each step signed by the tool that performed it. That chain is what "provenance" means in practice.
Newer versions of the specification add a soft binding: an invisible watermark or a perceptual fingerprint that can find the manifest again if a platform has stripped it from the file. OpenAI pairs its C2PA manifests with Google's SynthID watermark for this reason, and our SynthID guide explains how the two marks work together.
What "signed at capture" means
A manifest can be added at any point in an image's life, and the point matters. A manifest written by an editing application says what that application did; it knows nothing about the file's history before it was opened. A manifest written by the camera at the moment of exposure covers the whole history, because there is no earlier step.
That is why capture signing is the goal for anyone who needs to prove a photograph is a photograph. The camera holds a private key in a secure chip, signs the image before it leaves the device, and the record names the camera and the time. The Leica M11-P was the first production camera to do this, in late 2023. Sony, Canon and Nikon followed, and Google's Pixel 10 brought default-on capture signing to phones in 2025.
Capture signing proves that a specific device produced this exact file. It does not prove that the scene in front of the lens was real, a point we return to below.
What a missing manifest does and does not mean
Most images have no Content Credentials, and that will stay true for years. Only a short list of cameras sign at capture. Most editing software does not write manifests unless the feature is switched on. Instagram, Facebook and WhatsApp strip the metadata on upload, and screenshots never carry it. So a file without a manifest is the normal case, and its absence says nothing about whether the image is a camera photo or a generation.
The reverse misreading is just as common. The CR icon on an image does not mean the image is AI. It means a signed record exists, and that record is as likely to describe a Leica capture as a Firefly generation. What the record says is written inside it, in the assertions. Our article on why the icon is not an AI label walks through the misreading and how to avoid it.
How to read Content Credentials
You do not need special software to inspect a manifest. Drop the file into our free Content Credentials Inspector; it reads and validates the record in your browser without uploading the image. Adobe's Content Authenticity verify site does the same, and Google Search, Lens and Circle to Search expose the record through "About this image".
Whatever tool you use, look for three things. Who signed the record: a camera maker, a software vendor, a news organisation, an AI provider. What the record says happened: a capture, an edit, a generation. Whether the signature is still valid: an intact signature means the file has not changed since signing, and a failed one means it has.
A validator reports one of a few outcomes. Signed and intact with no edits recorded. Signed with edits listed. Generative AI declared in the manifest. Validation failed, because the bytes no longer match what was signed. Our step-by-step testing guide shows each outcome on real files and includes a known-good test image you can use to check that a validator reports the right result.
Who signs today
Adoption in 2026 is real but uneven, and the details decide what a manifest is worth in each case.
Cameras. Seventeen bodies from Leica, Sony, Canon and Nikon sign photos at capture, and the list grows with firmware. Fujifilm and Panasonic have not shipped it. The full list, model by model, is in our C2PA camera guide, and the camera check tool answers the question for a single model.
Phones. Google's Pixel 10 and 11 sign every photo from the camera app by default, with the key held in a hardware security chip. Samsung attaches credentials only to images edited with its AI tools. Apple does not use C2PA: the iPhone 18 Pro ships Apple's own Reference Image instead, readable only in Apple's apps. Our smartphone guide covers each vendor and what to do about the iPhone gap.
AI generators. OpenAI has attached manifests to DALL-E output since February 2024 and now adds SynthID as well. Adobe Firefly signs everything it produces. Midjourney and self-hosted Stable Diffusion ship nothing. The comparison is in our generator watermark table.
Editing software. Adobe Photoshop and Lightroom can read a manifest and extend the chain with their own signed edit entries. Most other editors drop the record on export.
Platforms. LinkedIn reads the record at upload and shows a Content Credentials label with a provenance panel. Meta reads it at upload and uses it only to label AI content. X strips it. The platform-by-platform guide lists who keeps, shows or removes the record.
Where C2PA stops
A valid manifest proves two things: which key signed the file, and that the bytes have not changed since. It proves nothing about the truth of the content, and three cases show the boundary.
A signed image of a screen. Point a signing camera at a monitor showing an AI image and the camera will sign the result correctly. The manifest says a Leica captured this file at this time, and that statement is true. The scene was still a monitor.
A signer fed manipulated content. In September 2025 a researcher used the Nikon Z6 III's multiple-exposure mode to make the camera sign an image it had not really captured. The signature was valid and the certificate chain checked out. Nikon suspended its authenticity service and invalidated every certificate it had issued. The full account is worth reading, because the cryptography worked exactly as designed and the flaw sat one level above it.
Edits outside the chain. A file edited in software that does not support C2PA and saved again loses the manifest or fails validation. The record speaks only for the steps it witnessed. Nothing in the standard can vouch for what happened in a tool that did not participate.
Trust in the signer is a fourth limit. A signature from any certificate validates mathematically, so validators compare the certificate against a trust list of known camera makers, software vendors and organisations. A file signed with an unknown or test certificate should be reported as valid but untrusted, and a good tool makes that distinction visible.
C2PA vs. AI detection
AI detectors and provenance records answer different questions, and the difference matters when a decision rests on the answer.
| AI image detection | Provenance (C2PA) | |
|---|---|---|
| Approach | Looks for statistical traces of generation in the pixels | Reads a signed record written when the file was created or edited |
| Output | A probability ("82% likely AI") | Facts: who signed, when, what actions were recorded |
| Reliability | Falls as generators improve; real photos get flagged | Stable; a signature is either valid or it is not |
| Coverage | Any image, including unmarked ones | Only files whose tools wrote a manifest |
| Failure mode | False positives on denoised or upscaled photos | Silence: most files have no record at all |
The two are complementary. A detector can screen an unmarked image; it cannot prove anything. A manifest can prove origin for the files that carry one; it cannot say anything about the rest. Our false-positive article covers what happens when a detector is treated as proof, and provenance vs. AI detection goes deeper into the comparison.
C2PA and the law
The EU AI Act's Article 50 has applied since August 2, 2026. It requires providers of generative AI to mark synthetic images in a machine-readable way, so that they can be detected as artificially generated. The regulation is technology-neutral and does not name C2PA, but a signed manifest that declares AI generation is the most direct way to meet the wording. Our EU AI Act article sets out the deadlines and penalties.
In courtrooms the standard plays a different role. A capture manifest is one piece of a chain of custody, alongside the RAW file, hashes and logs. Our chain of custody guide explains where a C2PA record fits among the other evidence.
How Lumethic uses C2PA
Lumethic works on both sides of the standard. The inspector reads any manifest a file carries. The verification service adds what a manifest cannot: a forensic comparison between a finished JPEG and the RAW file the camera wrote, which checks that the picture is consistent with a real capture rather than only that the file is unchanged since signing.
When that comparison passes, the result is written into a C2PA manifest attached to the image, with any earlier manifests preserved as ingredients. We call this verify, then sign. The signature then carries a claim the standard alone does not make: that the content was checked before the seal went on.
Frequently Asked Questions
What is the difference between C2PA and Content Credentials? C2PA is the technical standard. Content Credentials is the public name for a record that follows it, and the CR icon is its badge. A file with Content Credentials contains a C2PA manifest.
Does the CR icon mean an image is AI-generated? No. It means a signed provenance record is attached. That record may describe a camera capture, an edit or an AI generation; the icon itself carries no verdict. Open the record to see what it says.
Does a photo without Content Credentials come from AI? No. Most photos have no record, because only a small number of cameras and applications write one and most platforms strip it. Absence is the normal state and proves nothing either way.
Can a C2PA manifest be faked? The signature cannot be forged without the signer's private key, and any change to the file after signing breaks validation. What can happen is that a legitimate signer is fed content it should not sign, as in the Nikon Z6 III case, or that a file is signed with a certificate that no one trusts. Validators check certificates against a trust list for that reason.
Which cameras and phones sign photos with C2PA? Seventeen cameras from Leica, Sony, Canon and Nikon, plus Google's Pixel 10 and 11, sign at capture as of 2026. The camera guide and phone guide list every model.
How do I check a photo for Content Credentials? Drop it into the free Content Credentials Inspector. It reads and validates the manifest in your browser without uploading the file, and the testing guide explains each possible result.
