A photo, a video, or a document lands on your screen. No badge, no watermark, no manifest. Nothing tells you where it came from. If you’ve followed the last few years of AI news, the reflexive read is probably “no badge, probably fake.”
That read is wrong, and it cuts both ways: it gets genuine evidence thrown out of court and gets forgeries taken at face value. Content provenance records a file’s history, its origin, edits and tools. It does not certify that the content is true.
You’ll walk through what C2PA Content Credentials do, why a missing credential proves nothing, how the major systems compare, and how the liar’s dividend turns absence against real evidence. This is the corner of the synthetic-media trust gap that rarely gets argued head-on.
What is C2PA Content Credentials and how does content provenance actually work?
Content provenance is a file’s history: when and where it was made, what edits happened, and whether AI was involved. Content Credentials is the user-facing name for a C2PA Manifest, a cryptographically signed record of that history. It answers “is this what it claims to be?” at the moment of creation.
Keep two ideas apart. Provenance is history. Authenticity is whether the content truthfully represents reality. C2PA provides the first. As the standard’s own explainer is careful to say, Content Credentials don’t judge whether provenance data is true, only whether the record is well-formed, untampered and valid.
A C2PA-aware tool writes out assertions about the file and signs them with a private key. A SHA-256 hash hard-binds the manifest to the content, so any later change is evident. The binding is tamper-evident: edit the file and the hash won’t match.
The standard comes from the C2PA coalition, founded in 2021 by Adobe, Arm, the BBC, Intel, Microsoft and Truepic, with OpenAI, Google and Meta later joining. Adoption is opt-in, a detail that comes back to haunt the standard. Beside the signed manifest sit provider-specific watermarks, a softer signal that hints one model was involved. Provenance is the mirror image of deepfake detection: one verifies at creation, the other guesses after the fact.
Why does the absence of a content credential or watermark prove nothing about a file?
Here’s the asymmetry the argument turns on. A valid, verified credential is positive evidence of a file’s origin and chain of custody. A missing credential is the opposite: no information at all. Presence proves. Absence proves nothing.
Credentials disappear for innocent reasons. Platforms strip metadata to save bandwidth or protect privacy, and screenshots, recompression or format conversion do the same. A viral image can shed its provenance trace between publication and repost.
The distinction is between a broken chain and a chain that never existed. Tampering after signing is worth investigating; a chain that never existed tells you nothing. Neither proves forgery. As a journalist’s guide to C2PA puts it, most images online today have no manifest at all, and absence of a badge is “not evidence of fakery, it’s absence of data”.
Watermarks weaken the negative further. A watermark only answers a narrow question about one specific model. It can’t detect other models’ output, and it says nothing if the mark has been removed.
The result is a two-tier web: verified content on one side, everything else on the other. Most content in circulation is unlabelled by default, so treating absence as suspicion writes off the bulk of the internet by accident.
C2PA Content Credentials vs SynthID vs Meta Content Seal for provenance?
Once you accept that rule, the three systems become easier to compare, because they answer different questions. It comes down to interoperability versus fragmentation.
C2PA Content Credentials is the open standard: a signed manifest any tool can read and verify, but one whose coverage is limited by opt-in adoption. Because the signal sits in file metadata, it’s readable and can be stripped.
Google SynthID goes the other way, embedding an invisible watermark directly into the pixels that persists through screenshots, resizing, compression and colour adjustments. That durability separates it from a basic watermark, which strips away easily. SynthID interoperates with C2PA, and OpenAI has adopted it: watermarking survives transformation while metadata carries richer context.
Meta Content Seal sits at the opposite end. It’s proprietary, works only with Meta’s own Muse AI model, and doesn’t interoperate with anything else. Its July 2026 launch drew scepticism, with TechBuzz relaying The Verge’s doubt about the need for a single-vendor mark when open standards exist.
The same walled-garden logic applies to Anthropic’s Claude watermark, a SynthID-Text variant only checkable with Anthropic’s key. That’s the tension: open standards like C2PA and SynthID can be verified by anyone, while proprietary marks work only inside one vendor’s walls. It’s why the EU AI Act’s marking rules lean on interoperable, machine-readable marking, and why your team should question any vendor signalling claim that only the vendor can read.
What is the “liar’s dividend” and why does it matter in the provenance era?
The liar’s dividend is the term Bobby Chesney and Danielle Citron coined in 2019 for what happens when deepfakes make it easier for liars to dodge accountability for things that are actually true. As the public learns video and audio can be convincingly faked, bad actors get a free pass to dismiss genuine evidence as “probably AI”.
It is the presence/absence asymmetry’s downstream consequence. Because a missing credential carries no signal, doubt can be raised about anything, true or false. And because unverifiable media is now the norm, any genuine record can be waved away as unprovable.
In one 2026 case, a New York appeals court reversed a finding because third-party videos weren’t properly authenticated, despite details matching the scene. Newsrooms hesitate to report real events for fear the evidence is fake. Authentication burdens rise as fakery becomes more plausible, and awareness of detection’s limits feeds the doubt.
That’s why regulation exists. The EU AI Act’s Article 50 requires machine-readable marking of synthetic output, because voluntary standards leave a gap the law has to close.
That neutrality is the pressure point someone will exploit. Once unverified content is the default, “that’s probably AI” becomes a universal denial, and courts, newsrooms and boards carry the rising burden of proving what is real.
When a credential is missing, ask “what do we actually know?” Read a credential as a record of history. Read its absence as a gap in the data. And ‘nothing’ is a gap.
Frequently Asked Questions
Is content with valid Content Credentials automatically authentic?
No. A verified credential proves where a file came from and how it was edited, not whether its content is true. A photo can carry a perfectly valid C2PA Manifest and still be staged, mislabelled or misleading. The signature confirms integrity and origin; it says nothing about accuracy or intent. Treat the credential as a record of history, not a verdict on reality.
What is Google SynthID and how is it different from a normal watermark?
Google SynthID is an invisible watermark designed to survive editing, resizing, compression and screenshots while staying imperceptible. A basic watermark is easily stripped or cropped away, but SynthID is built to persist so detection still works at scale. It also interoperates with C2PA, and OpenAI has adopted it. Think of it as a durable, machine-readable signal rather than a visible stamp.
What is Anthropic’s invisible Claude watermark?
It is a provider-specific signal attached to Claude’s text output, based on a SynthID-Text variant. Because it belongs to one provider and does not interoperate across the ecosystem, it functions as a walled-garden mark rather than an open standard. It can help an informed verifier ask whether a specific model was involved, but its absence never proves that no AI was used.
Which organisations make up the C2PA coalition?
The C2PA was founded in 2021 by Adobe, Arm, the BBC, Intel, Microsoft and Truepic. OpenAI, Google and Meta have since joined the steering committee. The standard is also championed through Adobe’s Content Authenticity Initiative. Because adoption is opt-in, coverage across the wider web stays partial, which is a large part of why so much content circulates without any credential at all.
What is Meta Content Seal and why is everyone sceptical about it?
Meta Content Seal is a proprietary provenance signal that works only with Meta’s own Muse AI. It launched in July 2026 and drew immediate scepticism, notably from The Verge, because a single-vendor, non-interoperable mark fragments the ecosystem rather than strengthening it. It can confirm Meta’s involvement with its own model, but it cannot speak for any other tool or model.
Is open-source watermarking better than proprietary detection?
Open standards such as C2PA and SynthID generally serve the public interest better, because interoperable signals can be checked by anyone and across tools. Proprietary marks such as Meta Content Seal and the Claude watermark only work inside one vendor’s walls, so they fragment the ecosystem into a two-tier web. The EU AI Act’s machine-readable marking mandate pushes the field towards open, verifiable approaches.
Does stripping metadata mean a file has been tampered with?
No. Metadata disappears for entirely innocent reasons: screenshots, recompression, format conversion, re-saving and privacy-motivated stripping all remove credentials. A file with no manifest is consistent with genuine, unmanipulated content. The important distinction is between a broken chain, where a credential was removed or tampered with, and a chain that never existed. Only the former is suspicious, and even then it proves nothing on its own.
Can an AI provider simply remove its own watermark?
Yes, and that is one reason a watermark’s absence proves so little. A provider controls the signal it embeds, so it can withhold, weaken or strip it, and other providers’ models are invisible to it anyway. A watermark only ever answers a narrow question about one specific model. Its disappearance tells you nothing reliable about whether AI was involved.
What does “tamper-evident” actually mean?
Tamper-evident means any change to a file after signing becomes detectable, because the manifest is hard-bound to the content by a SHA-256 hash and a cryptographic signature. It does not mean the file cannot be changed. Someone can edit a file freely, but the alteration will break the hash and the verification will fail. That failure is a broken chain, which is different from a chain that was never present.
How do I check whether an image or video has Content Credentials?
Use a C2PA-aware tool, such as Adobe’s Verify site or software that displays Content Credentials, to read the manifest. The tool checks the signature against a trust list and confirms the hash still matches the file. If a manifest is present and valid, you learn the file’s origin and edit history. If none appears, you have learned nothing either way.
Will the EU AI Act marking rules apply to my content?
Article 50 requires providers and deployers of certain AI systems to mark synthetic output in a machine-readable way, and it applies to content reaching the EU market regardless of where it was made. If your AI-generated content is placed on the EU market, it likely falls within scope. The obligation exists precisely because the presence and absence asymmetry leaves a gap that voluntary standards cannot close.
What should a court or newsroom do when a file has no credential?
Treat the absence as a gap in the data, not as evidence of fakery. Instead of demanding a credential that may never have existed, ask what else can be corroborated: original device files, witnesses, chain of custody and independent forensic analysis. The practical trap is the liar’s dividend, where “that’s probably AI” is used to dismiss genuine evidence precisely because it carries no credential.
