Live-call verification protocol

7 things to ask on a suspicious video call.

Each one breaks a known real-time deepfake limit. Read in 10 seconds. Use on the call.

You're not being paranoid. Real-time deepfakes cannot handle these movements yet.

0
Challenges that exploit known real-time deepfake constraints Derived from published deepfake technical limitations (Rossler et al. 2019; Nguyen et al. 2022 survey)

The 7 Challenges

Tap any card to see what to watch for and why it works.

Say: "Turn your head quickly left, then right."
Watch: The jaw and ear region.
Why it breaks deepfakes: Real-time face-swap models struggle with large yaw-angle pose transitions; the swapped face may stutter, warp, or drop jaw alignment.
Technical constraint: Pose estimation failure at extreme angles
Say: "Hold up 3, then 5, then 2 fingers."
Watch: The fingers.
Why it breaks deepfakes: Real-time deepfake pipelines render faces well but hands poorly or not at all; fingers may blur, merge, or vanish.
Technical constraint: Hand-region rendering is not part of most face-swap models
Say: "Put your hand over the left half of your face, then move it away."
Watch: The boundary where hand meets face.
Why it breaks deepfakes: When an object crosses the face, landmark tracking loses points and the synthetic face glitches or fails to re-lock.
Technical constraint: Occlusion recovery lag
Say: "Turn and look over your right shoulder, then come back."
Watch: The profile and back-of-cheek area.
Why it breaks deepfakes: Profile and near-back-of-head views have minimal training data; the model produces artifacts or freezes.
Technical constraint: Sparse training data for non-frontal poses
Say: "Hold a cup or piece of paper in front of your face, then pull it away."
Watch: The face reappearing after the object passes.
Why it breaks deepfakes: Object-to-face boundary tracking breaks; the synthetic face may smear, delay, or snap back incorrectly.
Technical constraint: Dynamic occlusion segmentation
Say: "Cross your eyes for two seconds."
Watch: The eye region.
Why it breaks deepfakes: Unusual facial expressions are poorly represented in training data; the eye alignment distorts or the expression doesn't match.
Technical constraint: Expression mapping failure for non-prototypical expressions
Say: "Turn your chair all the way around so I can see the back of your head."
Watch: The back of the head.
Why it breaks deepfakes: Back-of-head has no face to track; real-time face-swap has nothing to render and either freezes the frame or shows a glitch.
Technical constraint: No facial landmarks to anchor the swap

How to use this on a call

Three steps. Takes 30 seconds.

1

Stay calm and keep the call going

Say: "Hey, can you do something for me real quick?" This sounds natural, not accusatory.

2

Pick any one challenge

Ask them to do it. You don't need all seven โ€” one or two is enough. Challenges 1, 2, and 3 are the fastest.

3

Watch the result

A real person does these instantly and naturally. A real-time deepfake will stutter, blur, freeze, or fail to render. If anything looks wrong, end the call and contact the person through a number you already know.

If the caller refuses or gets angry when you ask, that itself is a signal. A real loved one will just do it.

Save this. Share it with your family.

Deepfake scams target families โ€” especially older parents and grandparents. Save this page now so you have it before you need it.

This page works offline once loaded. No account, no tracking, no data collected.