The reliable tells are vanishing. Six-fingered hands and melting faces were a phase, not a permanent condition, and any guide built purely on visual glitches will be stale within months. What's worth learning is a sequence: look, then verify — and lean much harder on the second half.
First, look at the right things
Rendering still breaks in predictable spots, and it breaks worst at edges and in motion.
Watch what happens when a head turns to full profile. That's where synthetic video struggles hardest — an ear goes soft, a jawline drifts away from the neck, the arm of a pair of glasses sinks into skin instead of sitting on it. Jewellery is a great tell. Earrings and necklaces morph or briefly disappear as someone turns, because the model regenerates them every frame instead of tracking an actual object.
Then check texture. Hair moving as one solid mass instead of strands. Teeth rendered as a single white block with no separation between them. Skin polished to the point of looking waxy. Those are generation signatures, not lighting choices.
Pausing helps enormously. Step frame by frame through a moment of movement and look for a mask effect — skin tone or texture changing abruptly right along the edge of the face. Real footage has no seam there.
Then check the physics
Most people skip this part, and it's harder for a model to fake than a face is.
In a real scene, parallel lines — door frames, table edges, lane markings — all converge toward one vanishing point. Generated frames often contain two or three that disagree, which reads as wrong before you can explain why. Shadows behave the same way. If your subject is lit from the left but the chair beside them also throws its shadow leftward, something in that scene was invented.
Reflections, background text, and how objects meet surfaces deserve the same look. Models concentrate on the subject and get sloppy behind it.
Listen, too
People breathe while they talk, and they do it in predictable places — at clause breaks, before long sentences. Synthetic audio drops breath sounds in grammatically odd spots, or reuses the exact same breath several times over. Once you've caught a looping breath, you can't unhear it.
Now stop asking whether it looks real
Here's the shift that actually matters. Detection tools exist, and they check frame-level artefacts, audio inconsistencies, manipulation signatures. They're also losing an arms race, because every new generation of models gets trained specifically to beat the detectors that came before. Treating a tool's verdict as proof is a mistake.
The question that keeps working isn't "does this look real?" It's "can this be verified?" And that has an order to it.
Check the source. Who posted this, when, and do they have any history that predates the clip?
Reverse search it. If the footage is real and it matters, it exists somewhere else too.
Corroborate. Any event of consequence produces more than one video, from more than one angle, shot by more than one person.
Treat urgency as a red flag. Clips built to make you act right now — share before it's taken down, call this number, move this money — are betting on the gap between how long verification takes and how fast panic moves.
The one habit worth keeping
Most deepfakes travel as short cropped clips with no context around them, and that isn't laziness. The context is missing because the context would give them away.
So for anything that matters, go find the full video. If there isn't one, you've already learned the most important thing about the clip.
Image: Mikael Blomkvist, via Pexels
