Verification When Anything Can Be Fabricated

Discussion of synthetic media focuses almost entirely on one direction: fake material presented as real. That is a genuine problem and it is the smaller half.

The other direction is that real material can now be dismissed. When convincing fabrication is widely known to be cheap, anyone caught on record can claim the recording is fake, and the claim is no longer absurd. The law professors Bobby Chesney and Danielle Citron named this the liar's dividend in their 2019 work on deepfakes: the benefit that accrues to wrongdoers from a general climate of doubt.

Why that half is worse for the profession. Debunking a fabrication is difficult but tractable, because you are establishing a fact. Defending a genuine recording against a blanket accusation is much harder, because you are proving a negative to an audience that has been given permission not to believe you.

And the incentive structure is unfavourable. The fabricator needs only to create doubt. The journalist needs to establish confidence.

The practical consequence for how newsrooms work. Verification can no longer be something done to suspicious material. It has to be something built into how material is acquired, because the question is increasingly asked about your own reporting rather than about someone else's fake.

That reframing runs through this lesson. The goal is not detecting fakes. It is being able to demonstrate provenance for what you publish.

The instinct on encountering synthetic media is to reach for a detector, and it is worth understanding why that is the weakest available approach.

Detection is an adversarial problem. Any detector defines a target for generators to optimise against, and improvements in generation are continuous while detectors are trained on what already exists. That structural asymmetry means detector performance on the material that matters, the newest and best-crafted fakes, is systematically worse than benchmark figures suggest.

Second, detectors degrade badly on real-world material. A benchmark tests clean files. Actual journalistic material has been screenshotted, re-encoded, compressed by a messaging app, cropped and re-uploaded, and each of those transformations damages the signals detectors rely on.

Third, and most seriously for publication decisions, the errors run in both directions. A false positive means calling genuine material fake, which is a catastrophic error for a newsroom to publish. A false negative means passing a fabrication.

And there is a compounding problem. A detector output is a number, and numbers confer unearned authority. Ninety-two percent likely synthetic reads as a finding. It is an estimate from a model with unknown performance on this specific artefact.

So the practical position. A detector result is one weak signal among many, never a basis for publication either way, and it should never be reported to an audience as though it settled the question.

What actually works is older and less exciting, and it is the next step.

The methods that actually establish whether material is genuine are the ones the profession already had, and they are reporting techniques rather than technical ones.

Provenance. Who gave you this, and how did they get it? An unbroken chain back to a person who was there is worth more than any analysis of the file. Most fabrications fail here first, because the chain does not survive being asked about.

The original file. A screenshot of a video tells you almost nothing. The original file carries metadata, encoding characteristics and a resolution that a re-share does not, and asking for it is the single most productive request.

Corroboration. Is there a second source for the same event? A different angle, another witness, a document, a record?

Geolocation. Does the physical environment match where it claims to be? Buildings, signage, vegetation, road markings, the position of the sun. This is slow and it is decisive, because a fabricator has to get an entire scene consistent with a real place.

Chronolocation. Does the timing work? Weather records, shadow angles, what was under construction, who was present.

Internal consistency. Do the physics work? Does the audio match the room? Does the language match how this person speaks?

And directly asking. Contacting the apparent subject is verification, and it is skipped surprisingly often.

None of these are new. What has changed is that they have moved from being applied to doubtful material to being the routine standard for everything, because the cost of fabrication fell and the assumption of authenticity went with it.

How material should move from arrival to publication, with the decision points marked.

Material arrives. The first question is provenance: who provided it and how did they obtain it. An answer that does not survive follow-up questions is where most fabrications end, before any technical work.

Request the original file. A re-share, screenshot or forward is not the artefact; it is a copy with the useful evidence stripped.

Then corroborate independently. Geolocation against the physical environment, chronolocation against records, a second source for the same event. These are slow and they are what actually decides.

Detection tools sit off to the side deliberately, feeding in as one weak signal. They never determine the outcome in either direction.

Then the publication decision, and note that it has three outcomes rather than two. Publish as verified, with the basis stated. Publish with explicit uncertainty, describing what you established and what you could not. Or do not publish.

That middle option is the one under-used. Reporting that a video is circulating, that its origin cannot be established, and what checks were run is honest and useful. Presenting the same material as confirmed is where the damage happens.

There is a structural response to this problem that does not depend on detection, and it is worth understanding because it is where the industry is investing.

The approach inverts the question. Instead of trying to identify what is fake, attach verifiable information about origin to what is real. Content credentials record how a piece of media was created and what was done to it, cryptographically signed so that tampering breaks the signature.

The Coalition for Content Provenance and Authenticity, known as C2PA, is the main standards effort, with participation from camera manufacturers, software vendors and technology and media companies. The Content Authenticity Initiative is the associated adoption programme.

What this gives when it works. A signed record that an image came from a particular camera at a particular time, and that the edits applied were these. That is a positive claim about a real artefact rather than a probabilistic guess about a suspicious one.

The honest limitations, which matter for anyone relying on it.

Coverage. Only material captured and processed by participating tools carries credentials, which is a small fraction of what a newsroom receives, and almost none of what arrives from a witness with an ordinary phone.

Stripping. Metadata is routinely removed by social platforms, messaging apps and re-encoding, so absence of credentials proves nothing.

And it certifies capture, not truth. A signed photograph of a staged scene is a genuine photograph of a lie.

So provenance is a real improvement in the right direction and it does not remove the reporting work. It raises confidence where present and says nothing where absent, which is the correct shape for a signal but not for a gate.

One category deserves specific attention because it targets journalists directly rather than appearing in their coverage.

Voice cloning now requires very little source audio, and journalists have a great deal of audio in public: broadcast appearances, podcasts, conference talks, and voicemail greetings. That makes reporters unusually easy to impersonate.

The attacks that follow are practical rather than theoretical.

A call to a source, apparently from a journalist they trust, extracting information or an identity.

A call to a newsroom, apparently from a reporter in the field, requesting that something be published or changed under time pressure.

Fabricated audio of a reporter, used to discredit them.

And the reverse: a source contacting a journalist whose voice has been cloned from public material, feeding a false story.

What defeats this is not detection. It is out-of-band verification: confirming through a channel the attacker does not control.

The practical measures. Agree a verification method with regular sources in advance, before it is needed, which can be as simple as a callback to a known number or a shared word. Treat urgency as a warning sign, since manufactured time pressure is the common element in nearly all of these. And never let a voice alone authorise anything consequential, whether that is publication, payment or the disclosure of a source.

That last rule is the one that generalises. Voice is no longer an authentication factor. Anything that previously relied on recognising someone by how they sound now needs a second channel, and building that habit before an incident is far easier than after.

A practical problem with no clean answer, and worth thinking about before you are on deadline.

Covering a fabrication risks amplifying it. Many people who see your debunk will retain the claim and forget the correction, and the image or clip you republish to explain it will circulate detached from your explanation.

The research on corrections is more nuanced than the popular version. The strong claim that corrections reliably backfire and entrench belief has not held up well under replication; corrections generally do reduce false belief. But the weaker effect is robust: exposure to a claim increases its later familiarity, and familiarity is read as truth, so a debunk carries a real cost even when it works.

What follows practically.

Lead with what is true, not with the false claim. A headline that repeats the falsehood, even to deny it, is what most people will take away.

State the truth more prominently and more often than the claim being corrected.

Show the fabrication only when necessary to explain it, marked unmistakably, and consider whether describing it serves as well as showing it.

Give the audience the mechanism rather than only the verdict. Explaining how you established it was fabricated is more durable than asserting that it was, and it transfers to the next case.

And apply a threshold. Not every fabrication warrants coverage. Reporting on one with negligible reach can hand it an audience it never earned, and that judgement, about whether covering something serves readers or serves the fabricator, is exactly the kind of editorial decision lesson one said does not delegate.

Pulling three lessons into what someone doing this work should carry.

The compression is real and it sits in production, transcription, document handling and line editing. Those are large, unglamorous shares of the week, and taking them back is a straightforward gain with no editorial cost.

The one disqualifying risk is fabrication, and it has a bright line rather than a judgement call. Nothing inside quotation marks comes from a model. No fact reaches print unverified against a source. Citations are checked or they do not run.

The safe uses are identifiable by direction. Where you supply the material and the model reshapes it, invention is impossible. Where you ask the model to supply material, it is always possible and never visible.

Voice is mostly judgement rather than prose style, so it survives using the tool as an editor and erodes using it as a ghostwriter. And composition is where thinking gets checked, which is why keeping it is worth more than the time it costs.

Verification has become routine rather than exceptional, and it is reporting work: provenance, original files, corroboration, geolocation. Detection tools are a weak signal and provenance standards raise confidence where present without proving anything where absent.

And the liar's dividend means the harder task is increasingly defending what is real rather than exposing what is fake, which is an argument for being able to show your work.

The unifying observation. Every standard in this cursus predates the technology. What changed is that the assumption of authenticity, which the profession quietly relied on, is gone.