1. What counts as synthetic media
Synthetic media covers material generated or substantially altered by a model: a face swapped onto another body, a voice cloned from a short sample, a photograph produced from a text description, a video of an event that never took place. It shades into ordinary editing, which is why a binary real-or-fake question is usually the wrong one. Colour correction, cropping and noise reduction are unremarkable; removing a person from a scene is not. The operative question for a desk is whether the alteration changes what a reasonable viewer would understand the material to record. That framing also survives the next technique, which the current detection tooling will not.
2. How the material is produced, in outline
Face replacement systems learn a mapping between two sets of images and re-render one identity onto another’s movements. Voice cloning learns a speaker’s timbre and prosody from a sample that can now be very short, then reads arbitrary text in that voice. Diffusion models generate images from text by iteratively removing noise under the guidance of a learned representation. Understanding this at outline level matters editorially, because it tells you where errors are likely: hands and teeth in generated images, breathing and room tone in cloned audio, physics and reflections in generated video. It also tells you why those weaknesses keep disappearing as systems improve.
3. Provenance first, artefacts second
The most reliable verification route is not forensic but archival: establish where the file came from. Ask for the original file rather than a re-encoded copy, check embedded metadata while remembering that it is trivially editable and routinely stripped by platforms, and look for content credentials where the capture device or software supports them. Provenance standards such as C2PA attach a signed record of origin and subsequent edits to a file; where present and intact, that record is strong evidence, and where absent it proves nothing either way, because most genuine material carries no credential at all.
4. Detectors and their limits
Automated detectors report a probability, not a verdict, and their accuracy falls sharply on material that has been compressed, cropped, re-recorded from a screen or produced by a technique that postdates their training data. False positives on genuine footage are common enough to be dangerous, because a desk that publishes “analysis suggests this is fake” about real material causes a serious and hard-to-retract harm. Treat a detector score as one input alongside provenance, source interview and physical corroboration, and never quote a score to a reader without explaining what the tool measures and where it is unreliable.
5. Corroboration in the physical world
Synthetic material is weakest against the world it claims to depict. Identify buildings, signage, vehicle plates, road markings, vegetation and skyline against mapping and street-level imagery. Check shadow direction and length against the claimed time and latitude. Compare weather in the frame against records for that date and place. Look for other recordings of the same moment from different positions, and for people who would necessarily have been present. This work is slower than running a detector and far more conclusive, and it is the same discipline that verification desks have applied to user-submitted material for over a decade.
6. Triage under deadline
A suspect clip usually arrives when it is already spreading, so the desk needs a sequence it can run in twenty minutes rather than a research project. Establish the earliest available copy and who posted it. Contact that account. Ask what device captured it and request the original. Run a reverse image search on several frames. Check two physical details against independent records. Consult an expert if the material is consequential. If the sequence does not resolve, the honest output is a report about a circulating clip whose origin has not been established — clearly framed as such, and without embedding the clip in a way that amplifies it further.
7. The liar’s dividend
The wider harm of synthetic media is not only that false things are believed but that true things can be denied. Once audiences know that convincing fakes exist, any authentic recording can be dismissed as fabricated, and the burden of proof shifts onto the publisher. Newsrooms reduce that dividend by documenting provenance as a matter of routine on their own material: recording who captured what, when and on which device, retaining originals, and publishing that chain when authenticity is challenged. A publication that can show its working when questioned is far harder to wave away than one that simply insists.
8. Rules and disclosure duties
The EU AI Act introduces transparency obligations for certain synthetic outputs, including duties around machine-readable marking of AI-generated content and disclosure where material resembling real people, places or events could mislead, with specific consideration given to editorial contexts. National frameworks apply alongside it: in Ireland, Coimisiún na Meán oversees aspects of online safety and media regulation, and the Press Council of Ireland and the Office of the Press Ombudsman handle complaints against member publications under the Code of Practice. Obligations depend on role and timing, so read the current text and take qualified advice rather than relying on a summary.
Verification signals and how much weight they carry
| Signal | How obtained | Evidential strength | Failure mode |
|---|---|---|---|
| Original file from the capturer | Direct contact | High | Cannot always be obtained under deadline |
| Intact content credential | Embedded provenance record | High where present | Absent from most genuine material |
| Embedded metadata | File inspection | Low to medium | Editable and routinely stripped in sharing |
| Physical corroboration | Mapping, weather and imagery records | High | Slow; needs identifiable features in frame |
| Second recording of the same moment | Search and appeals | High | Often does not exist |
| Automated detector score | Detection tool | Low on its own | Degrades on compressed or novel material |
| Visual artefacts | Close inspection | Low and falling | Improves out of existence with each model generation |
Weight is cumulative and directional: strong provenance plus physical corroboration can support publication, while a detector score alone never should.
Rapid triage for a suspect clip
- The earliest available upload has been located, not just the most viewed copy.
- The apparent capturer has been contacted and asked for the original file.
- Reverse image searches have been run on at least three separate frames.
- Two physical details have been checked against independent records.
- Shadow direction and weather have been tested against the claimed time and place.
- Any detector output is recorded with the tool name, version and date.
- An expert has been consulted where the material is consequential.
- The published framing states what has been established and what has not.
- The clip is described rather than embedded where amplification is a concern.
If the sequence cannot be completed, publishing the uncertainty is a legitimate outcome. Publishing a guess dressed as a finding is not.
Questions readers ask
Can a detection tool prove a video is fake?
No. Detectors output probabilities that degrade on compressed, cropped or re-recorded material and on techniques newer than their training data. Use them alongside provenance and physical corroboration, never instead of them.
What are content credentials?
A signed provenance record attached to a file by supporting capture devices or software, recording origin and subsequent edits. Where present and intact they are strong evidence; their absence proves nothing, since most files carry none.
Should a newsroom publish a suspected deepfake at all?
Sometimes the circulation itself is the story. Where it is, describe the material and the state of verification rather than presenting it as a record of events, and consider whether embedding it adds anything beyond reach.