TWO THINGS TO JUDGE
A digital image has usually asked us to judge the image itself.
Is this what it claims to be? Where did it come from? Has it been altered? Is its source credible?
Provenance systems introduce something else to inspect. Media can carry signed claims about parts of its own history: how it was captured or generated, what tools acted on it, which changes were recorded, and which system signed those claims.
This is already more than a standards proposal. Provenance is present at different points in today's media stack, including capture, generation, platform processing and distribution. What remains unclear is how prevalent it is across the media people actually encounter, or how reliably those records survive from creation to consumption.
So the important change is not that media has suddenly become authenticated. It has not.
It is that, when provenance exists and remains intact, there can be two related things to judge: the media object and the record making claims about its history.
The second cannot tell us whether to believe the first.
It can give us more evidence with which to decide.
WHAT THE RECORD KNOWS
“Authenticity” makes this harder to understand because it compresses several different questions into one word.
A provenance record can contain signed assertions about capture, creation, generation, edits, transformations, prior assets and processing history. It can make parts of that claimed history inspectable and make tampering with the signed record detectable.
That can tell us, for example, that a particular system says an image was generated rather than captured, that an editing action appears in its history, or that another asset was used in its creation.
But these are statements about provenance, not about the truth of what the media communicates.
Authentic footage can carry a false caption. An unaltered photograph can be placed in deceptive context. A generated image can carry accurate provenance saying that it was generated. A recognised institution can still publish something wrong.
C2PA's own model preserves this distinction. Technical validation checks the relationship between an asset and the provenance claims attached to it. It does not turn that result into a judgement that the media is accurate, trustworthy or benign.
The record can tell us more about the media's claimed history.
Reality remains outside it.
WHO SPEAKS FOR THE RECORD?
Evidence has a source too.
In C2PA's core model, the signer behind a provenance assertion may be an application, a service, a device or another hardware or software product. Human and organisational identities can be attached through additional mechanisms, but they are not inherent in every credential.
So the existence of a signer does not automatically answer a more human question: who is responsible for this?
A camera can attest to aspects of capture without being responsible for what happened in front of it. Software can record that it performed an edit without becoming the author of the message created by that edit. A technically valid signature can identify the system making an assertion without telling us whether the person or institution using that system deserves trust.
This is where provenance begins to look more like evidence in the ordinary sense: the evidence itself has to be evaluated.
Which system made the claim? Which identity stands behind it, if any? Which platform preserved or displayed it? What significance should any of that carry?
An open standard can make claims interoperable without making those questions disappear.
Nor does this establish that provenance is moving society's trust from one set of institutions to another. We do not have that evidence.
It means only that a provenance record cannot interpret its own authority.
WHEN THE SECOND OBJECT DISAPPEARS
The second object can disappear even when the media remains.
Provenance can be lost when media passes through tools or services that do not preserve it. Metadata can be stripped. Files can be re-encoded. Screenshots and re-recordings can create new media objects without carrying the original credential forward.
C2PA includes mechanisms intended to make records more resilient or recoverable, and some distribution infrastructure can preserve provenance through transformations. Disappearance is not inevitable.
But neither is persistence assured.
The consequence is asymmetric.
Presence can provide positive evidence. Absence does not provide the inverse proof.
An image without provenance is not therefore synthetic. A video whose provenance chain has broken is not therefore manipulated. There are too many ordinary ways for legitimate media to arrive without an intact record.
That distinction matters because interfaces can encourage the opposite intuition. Research has found that incomplete or invalid provenance can sometimes make people distrust media that is otherwise honest.
A system intended to add evidence becomes misleading if the absence of that evidence is treated as evidence itself.
Provenance is strongest when it can say: here is a record you can inspect.
It is much weaker when asked to say: no record exists, therefore something must be wrong.
WHO NEEDS TO READ IT?
The familiar image of provenance is a visible label: an icon or disclosure that somebody notices and uses when deciding what to believe.
But a provenance record does not have to be read by an ordinary user to be processed.
Platforms can consume it automatically. Distribution systems can preserve it. Verification tools can expose it to professionals. User interfaces can choose to translate some of it into a visible disclosure.
These uses should not be collapsed into a single idea of adoption.
Current evidence shows that machine-readable provenance is being processed in deployed systems. It does not show that this has already produced better moderation at scale, faster newsroom verification, reduced fraud, lower brand-safety risk or improved rights outcomes.
The user evidence is similarly bounded.
A 2026 experiment involving news imagery found that showing C2PA provenance information increased perceived transparency and credibility and increased trust in the presented news source. Earlier research found a different problem: people could confuse the credibility of provenance with the credibility of the media itself, while incomplete or invalid provenance could cause honest media to be viewed with greater suspicion.
Broader research on synthetic-media labels also suggests that changing perceptions does not automatically change sharing or engagement.
The evidence therefore supports neither “people trust provenance” nor “people ignore it.”
A machine being able to read provenance is one outcome. A person understanding it is another. That person's behaviour changing is another still.
Provenance may eventually prove more consequential inside systems than in the iconography shown to users. Or its institutional benefits may prove modest. Current deployment lets us ask that question seriously; current evidence does not settle it.
THE RECORD CANNOT CLOSE THE CASE
Provenance may never become universal infrastructure.
It may remain one verification signal among many: alongside source reputation, reporting, watermarking, disclosure and ordinary human scrutiny. Cross-stack deployment shows that its function is real. It does not establish its eventual scale.
What makes provenance distinctive is narrower than dominance and more useful than a promise of certainty.
Its distinctive function is to associate signed claims with aspects of a media object's history. Those claims can become available for inspection without becoming an automated verdict about whether the object is true, accurate or trustworthy.
That changes what is available for judgement without automating judgement itself.
The record may tell us that a system captured an image, that software changed it or that a particular signer made an assertion. We still have to decide whether the record is sufficiently complete, whether its source matters, whether the context is honest and whether the claim made through the media is true.
The important addition is not a verdict.
It is another object of scrutiny.
A piece of media can arrive with more of its history open to challenge. Whether that history makes the claim believable remains a judgement outside the record.