How to Spot a Deepfake: Face-Swap Tells That Still Work
Deepfakes put a real person's face on someone else's video. These are the visual and behavioral tells that still expose them — and what to do when you can't tell.
A deepfake hijacks the most trusted signal humans have — a familiar face. That’s why deepfakes power the worst abuse: fake celebrity endorsements, executive fraud on video calls, political fabrications, and non-consensual imagery. The tells below focus on where face-swaps still fail.
Where face-swaps break
- The face boundary. The swap is a patch applied to a head. Look at the jawline, hairline, and ears — flickering, blur, or a skin-tone seam where the patch meets the original.
- Profile views. Ask for (or wait for) a full side view. Many face-swap models degrade sharply past 60 degrees of head rotation.
- Occlusion stress-tests. A hand passing over the face, glasses coming off, drinking from a cup — objects crossing the face boundary cause smearing or ghost frames.
- Blinking and micro-expressions. Unnatural blink rates, eyes that don’t crinkle with a smile, a forehead that stays smooth while eyebrows move.
- Teeth and tongue. Interior mouth detail is hard; watch for a dark, undefined mouth cavity or teeth that look like a single white strip during speech.
- Lighting mismatch. The swapped face is lit by the source footage’s lighting. If the room’s light comes from the left but the face is lit frontally, that’s a seam you can’t unsee.
Behavioral tells (for live calls and “video messages”)
- The person avoids turning their head or keeps an unnaturally fixed pose.
- Latency between voice and lips, especially at the start of sentences.
- Refusal of spontaneous requests — “wave your hand in front of your face”, “stand up and turn around”. Fraudsters end the call instead.
- Urgency plus secrecy. Every deepfake fraud script combines “act now” with “don’t tell anyone”. The social engineering is the constant; the technology just improved.
The verification mindset
For anything involving money, credentials, or reputation: verify through a second channel you already trust — call the person back on their known number, check the company’s official site, search for the “announcement” from reputable outlets. A deepfake defeats your eyes; it doesn’t defeat a callback.
When your eyes aren’t enough
Face-manipulation detection is one of the specific things automated analysis does better than humans: models are trained on the boundary artifacts, blending signatures, and frequency-domain fingerprints that face-swaps leave behind.
Frequently asked questions
What's the difference between a deepfake and an AI-generated video?
A deepfake typically maps a real person's face onto existing footage of someone else, while an AI-generated video is synthesized from scratch by a text-to-video model. Deepfakes are the tool of choice for impersonation because they hijack a real, recognizable identity.
Can deepfakes work in live video calls?
Yes. Real-time face-swap tools run during video calls, which is how several high-profile corporate frauds were executed. Asking the person to turn fully sideways or pass a hand in front of their face still breaks many live face-swaps.
Are deepfakes illegal?
Depends on jurisdiction and use. Many places now criminalize non-consensual intimate deepfakes and election-related synthetic media, but a large gray zone remains. Platform rules are typically stricter than the law.