How do I verify a viral video is real before my newsroom runs it?

From the trust and safety side, where we see thousands of these a day rather than one on deadline. Our order of operations is provenance first, content second. Where did this file first appear, who posted it, does that account have a history, and can we find any earlier copy? Half of what gets flagged to us as a deepfake turns out to be a real clip from a different year and place with a new caption. Only after that do we look at the pixels.

We run automated detectors, and I will be honest about how we use them: as one input into a queue priority, never as a decision. Their outputs move with compression, re-encoding and resolution, and a messaging-app re-upload is the worst case for all three. For a newsroom I would treat them the same way. A detector score should change what you check next, not what you publish.

Agree with Ingrid, and one framing that may help the desk: detection is adversarial. Any specific tell, blinking, teeth, hairline, becomes a training target for the next generation of tools the moment it is published. A detector is a snapshot of what last year's fakes got wrong. So any process that depends on a tell or a tool has a short shelf life, and any process that depends on provenance and corroboration does not.

Also, be as sceptical of a "real" verdict as a "fake" one. A clean result on a heavily re-encoded file mostly means the detector could not see anything, not that there was nothing to see.

@Tomas Reyes this is exactly what I was after, thank you. We already use the plugin for stills and I had not thought to run the keyframes through the same path we use for photos. Writing the checklist up for the desk now, with "earliest copy" at the top and "detector result" near the bottom with a note that it is a lead, not a clearance.

audio person here so a tip from my side: listen on headphones for room tone. real recordings have the room in them, air conditioning hum, chairs, someone coughing. a cloned voice is usually clean and dead behind the words. if there is background noise under the speech the audio is real, you can move on and worry about the picture

One step I would add to the written checklist, since it is the one that gets skipped under deadline pressure: the rights and legal position of the clip itself. Who owns it, do you have permission to republish, and if it depicts a private event, what is your basis for running it at all? That matters whether or not it is genuine. If it is a fabrication, you also have a separate exposure for republishing it, which the thread on the legality of deepfakes touches on. A verification checklist that ends at "is it real" is half a checklist.

Everything above is right. I will add the boring step that still works. Call the person in the video. Call their office. Call someone who was at the event. Three phone calls take twenty minutes and beat every tool in this thread. If nobody will confirm it and nobody will deny it, that is information too. Hold it.

@Jules the first half is a useful listening habit, but the conclusion does not follow. Background noise is trivial to add after the fact: generate the voice, mix in a recording of a conference room, re-encode once, and the result has all the room you are describing. Lip sync tools also work from an audio track and can take whatever audio you give them. So noise under the speech tells you nothing about whether the words are genuine. It is the same lesson as the blinking tell in the chapter Tomas linked: as soon as a feature is known to be used as a check, it becomes a feature people make sure to include. The voice cloning thread has the fuller discussion.