Stop trying to spot it by eye
The most common advice about deepfakes — look for weird blinking, count the fingers, watch the teeth — is quietly becoming obsolete. Those tells catch clumsy fakes, but the technology improves every month, and the fakes that actually matter in elections have already outrun them. The clearest proof is that the most effective political deepfake so far was audio, not video, and audio has no fingers, teeth, or blinking to inspect at all.
So start by dropping the assumption that you can win a staring contest with a modern fake. You usually cannot, and trying to teaches false confidence. The good news is that you rarely need to, because a fake's weakest point is almost never the media itself — it is everything around the media. That is the same lesson as how to fact-check any claim: judge the source and the context, not just the artefact.
Provenance beats pixels
The reliable question is not "does this look real?" but "where did this come from?" A deepfake can perfectly fake a face and a voice; what it cannot fake is a credible, traceable origin. Real events leave trails — a full recording, a press event, coverage by outlets that would stake their reputation on it. A fake usually appears as a short, cropped, low-quality fragment from an account you have never heard of, with no original anywhere behind it.
Checking provenance is unglamorous — searching for the earliest version, looking for a primary source, seeing who else reports it — but it is what actually works, and it works regardless of how convincing the pixels or the waveform are.
The timing trap
Deepfakes in elections are weaponised less by quality than by timing. The playbook is to release a fake late — ideally in the final days before a vote, when there is no time to debunk it and, in some countries, a campaign-silence period legally muzzles the people best placed to respond.
Slovakia's 2023 parliamentary election is the reference case. Roughly 48 hours before polls opened, fabricated audio purporting to catch a party leader and a journalist discussing vote-rigging spread during the pre-election moratorium, when rebuttal was constrained. The party went on to lose narrowly. The honest lesson is not that the fake decided the election — that is unprovable — but that its timing was the entire design. So a practical rule follows: be most suspicious of a clip precisely when it feels most damning and lands at the most decisive moment.
The visual and audio tells (and why they're fading)
Used carefully, the old signs still have some value against weak fakes:
| Medium | Possible tell | Caveat |
|---|---|---|
| Video | Unnatural blinking, odd hands/teeth, blur at the hairline, lip-sync drift | Vanishing in good fakes |
| Video | Lighting or shadows that don't match the scene | Easy to get right now |
| Audio | Flat emotion, odd breathing, robotic cadence | Often masked by low quality |
| Any | Too short, too cropped, no context | The most durable signal |
Two warnings. First, the presence of a tell suggests a fake, but the absence of tells proves nothing — the best fakes have none. Second, do not fall for the mirror-image trap: the existence of deepfakes lets people dismiss genuine recordings as fake. That reflex — the liar's dividend — means "it's probably a deepfake" is not a free pass either, for you or for the person caught on tape.
Tools help but don't decide
Automated detectors exist, and organisations such as WITNESS have pushed for wider access to them. They are worth using — but as a hint, not a verdict. Detection consistently lags generation, and the tools produce false positives and false negatives, so a single tool's output should sit alongside provenance and context, never replace them.
The durable defences are the human ones covered in deepfakes in elections: prebunking so audiences expect last-minute fakes, fast trusted rebuttal, and a habit of asking "where did this come from, and why now?" before "does this look real?" Get that order right and you are already ahead of the fakes — without having to out-stare a single pixel.