Verifying synthetic media

1. Ask a better question than “is it fake?”

Almost all published media has been altered in some way — cropped, colour-corrected, compressed, stabilised, captioned. A binary authenticity test therefore fails at both ends: it flags ordinary editing and it misses a convincing fabrication that has been re-encoded enough to defeat analysis. The question that survives contact with real material is narrower and more useful: does the alteration change what a reasonable viewer would understand this file to record?

That framing produces a workable threshold. Noise reduction does not change the record. Removing a person from a frame does. A cloned voice reading a statement the speaker never made does, absolutely. It also survives technological change, because it is a question about meaning rather than about the artefacts of any particular generation method — and artefacts are precisely what each new model generation eliminates.

2. Provenance before forensics

The strongest evidence about a file is almost never inside the pixels. It is the chain from capture to receipt. Ask who recorded it, on what device, where they were standing, what happened immediately before and after, and whether anyone else was present. Request the original file rather than a shared copy, because platforms re-encode uploads and strip metadata, destroying exactly the information verification depends on.

Where provenance standards are in use, a signed record of origin and subsequent edits may travel with the file. Such credentials are strong evidence where present and intact, and neutral where absent, since the overwhelming majority of genuine material carries none. Treat embedded metadata with more caution still: it is informative, trivially editable, and routinely removed in ordinary sharing, so it can support a conclusion but should never carry one alone.

3. Source contact and chain of custody

Contacting the apparent capturer resolves more cases than any technical step, and it is skipped most often, because it is slow and sometimes unsuccessful. Ask questions whose answers can be checked against the frame: which direction were you facing, what was behind you, what could you hear, what time did you leave. A person who genuinely recorded a scene answers these easily and consistently; someone recirculating material usually cannot.

Record the chain as you go: who supplied the file, when, in what form, and what was asked and answered. This record has two uses. It supports the published account if the material is later challenged, and it prevents the common failure in which a clip is verified by one desk, passed to another, and gradually loses the qualifications that were attached to it. Chain notes should travel with the asset in the content system, not live in a personal inbox.

4. Physical corroboration

Synthetic material is weakest against the specific, checkable world. Identify buildings, signage, road markings, vehicle types and registration formats, vegetation, and the skyline. Compare them against mapping and street-level imagery, remembering that such imagery has its own capture date and may be years old. Check the weather in frame against records for the claimed date and location. Look for other recordings of the same moment from a different position, which is the single most conclusive corroboration available.

5. Chronolocation arithmetic

Shadows are one of the few things in a frame that obey arithmetic. If a vertical object of known or estimable height casts a measurable shadow on level ground, the sun’s altitude follows directly, and altitude constrains the possible times for a given date and latitude. It is not a precise clock, but it is often enough to eliminate a claimed time entirely — which is usually what a verification desk needs.

The third line matters for a different reason. Clips are frequently described as showing a continuous sequence when the frame count and stated frame rate do not match the claimed duration, which is a strong indication of cutting. Both calculations are cheap, repeatable and independent of any detection tool.

6. Detectors: how to use a probability

Automated detectors output a likelihood, and that likelihood is conditional on material resembling their training data. Compression, cropping, re-recording from a screen and any generation technique newer than the model’s training set all degrade performance, often severely. False positives on genuine footage are common enough to be a serious hazard: publicly casting doubt on authentic material causes a harm that is very difficult to undo.

Use detectors as one input, never as a verdict. Record the tool name, version and date alongside the score, because a score without them is uninterpretable later. If a score is mentioned to readers at all, explain what the tool measures and where it fails; a bare percentage reads as certainty and will be quoted as one. Where the material is consequential, consult a specialist rather than relying on a self-service tool.

7. Publishing when verification is incomplete

Deadlines do not wait for certainty, and sometimes the circulation of a clip is itself the story. The honest output in that case is a report about a circulating file whose origin has not been established, framed as such from the headline onwards, and stating in plain terms what has been checked, what could not be checked and what would change the assessment.

Presentation decisions matter as much as wording. Consider describing rather than embedding, so the item does not become a distribution channel for the material. Avoid using a frame from the clip as the lead image, which visually asserts what the text is careful not to assert. And commit to updating the item when more is learned, with a visible note rather than a silent edit.

8. Corrections, takedowns and the archive

If verification later fails, the correction has to be at least as visible as the original. Amend the item in place, retain and clearly mark the superseded wording, state what was wrong and when it was discovered, and update any social posts that carried the claim. Deleting the item outright is usually the worst option: it removes the record without removing the belief, and it looks like concealment.

Requests to remove material raise separate questions — legal exposure, the privacy of identifiable individuals, the safety of people depicted, and the public interest in a durable record. These do not resolve on general principles and are jurisdiction-specific. Handle them under a written policy and take advice from a qualified professional on individual cases.

9. The liar’s dividend and your own material

As convincing fabrication becomes ordinary, denial becomes cheaper: any authentic recording can be dismissed as synthetic, and the burden falls on the publisher to prove otherwise. The defence is documentation applied to your own output. Record who captured what, when, on which device; retain originals in an archive that preserves them unaltered; keep the chain notes; and be prepared to publish that chain when authenticity is questioned.

This is unglamorous work with no visible benefit until the day it is needed. It is also the only credible answer to a well-organised denial, and it is far easier to adopt as routine practice than to reconstruct in the middle of a dispute.

10. The verification chain, in order

Run the steps in sequence and stop when the material is either established or established as unestablishable. The order matters: the cheap, conclusive steps come first, and the expensive, inconclusive ones come last.

1. Step 1 — locate the earliest copy Search for prior uploads rather than accepting the most-viewed version as the original.
2. Step 2 — contact the apparent capturer Ask checkable questions about direction, surroundings, sound and timing; request the original file.
3. Step 3 — inspect provenance Check for content credentials and embedded metadata, recording what is present and what is missing.
4. Step 4 — corroborate physically Match landmarks, signage and vegetation; test shadows and weather against the claimed date and place.
5. Step 5 — search for a second recording Look for the same moment captured from another position or by another person.
6. Step 6 — run detectors, with caution Record tool, version, date and score; treat the result as one input among several.
7. Step 7 — consult a specialist For consequential material, obtain an expert assessment before publication.
8. Step 8 — state the position Publish what is established, what is not, and what would change the assessment.

Verification checklist for suspect material

• The earliest available copy has been located, not the most-shared one.
• The apparent capturer has been contacted and asked checkable questions.
• The original file has been requested, and its absence noted if not supplied.
• Content credentials and metadata have been inspected and the findings recorded.
• At least two physical details have been matched against independent records.
• Shadow and weather consistency has been tested against the claimed time and place.
• A search for a second recording of the same moment has been carried out.
• Any detector output records the tool, version, date and score.
• The chain of custody note is stored with the asset, not in a personal inbox.
• The published framing distinguishes what is established from what is not.
• A specialist has been consulted where the material is consequential.

A completed checklist is not a guarantee of authenticity. It is a record of the work done, which is what allows a publication to explain its reasoning if the assessment is later challenged.

Questions readers ask