How to spot a deepfake or AI-generated video

Deepfake video no longer looks like the glitchy, uncanny-valley clips from a few years ago. Current models produce faces that blink naturally, track light correctly, and move with plausible physics. The obvious tells are mostly gone. What's left are subtler signals — ones that require watching for pattern rather than glancing for glitches.

Start with the eyes and edges. Reflections in the eyes should match the surrounding light source; synthetic faces often show reflections that are slightly too uniform or don't shift as the head turns. Look at the boundary where hair meets forehead, or where glasses meet skin — these transition zones are still where generation models most often show softening, blurring, or faint warping under close inspection. Ears, in particular, remain a weak point: earrings, piercings, and the fine cartilage detail of a real ear are difficult for most models to render consistently across frames.

Audio-visual sync is another reliable signal. Watch the mouth shape against the sound — plosive consonants like "p" and "b" require a specific mouth closure that generated video sometimes renders a frame or two late or too soft. Background consistency matters too: does the lighting on the subject's face match the lighting in the room behind them? Does a moving object in the background (a fan, a flag, foliage) move with natural randomness, or does it loop in a way that feels too smooth?

None of these signals are conclusive alone, and that's the honest answer — visual detection is becoming a losing game for the unaided eye. That's part of why AI-detection tools analyse statistical patterns invisible to human viewers rather than relying on visual glitches at all.

Lenstrum's AI Involvement tab flags AI-generation probability for any video URL. Try it free →