What liveness tests still work against deepfake video calls in 2026?

Tests that still work: asking the person to hold a specific physical object you name in the moment (deepfakes cannot synthesise object interaction with the face reliably); asking them to move to a different room and describe it in real time; requesting they hold their hand flat against their cheek and rotate slowly (occlusion of the face by a real hand degrades the deepfake rendering significantly); and asking them to read a sentence you type during the call, while watching for lip-sync micro-errors at the syllable level.

Why do standard blinking and head-turn tests no longer detect deepfakes?

Standard liveness tests were designed to detect static image spoofing and early deepfakes. Modern face-swap models are trained on thousands of hours of video including blinking, smiling, and head movement, so they reproduce these motions naturally. The weakness of current models is not in smooth motion — it is in handling unexpected, compound physical interactions that were not part of their training distribution.

If someone passes a video call liveness test, are they real?

Not necessarily. Passing a video call test means the technology is not yet good enough to fail your specific test — it does not confirm identity. The only confirmation of identity is verification against public records. A real person who exists in registries, employment databases, and social systems is verifiable regardless of what any video call shows.

AI Fraud

Liveness Check Failures: When Video Calls Are Faked

By Michael Muinov, Founder and Lead Investigator · Active casework since 1999 · Updated 20 May 2026

A video call used to be the gold standard for proving someone is real. In 2026, real-time deepfake injection means it no longer is. Here is what still works — and what does not.

Quick answer

Can a video call be faked to look completely real?

Yes. Real-time face-swap technology can replace a scammer's face with a synthetic or stolen identity on a live video call with latency low enough to avoid obvious lag. This technology is available in commercial tools and does not require specialist hardware. Standard liveness checks — blinking, nodding, turning your head — are now easily defeated because these motions are part of the training data for every major deepfake model.

The good news is that current real-time deepfakes have predictable failure modes. Tests that exploit those failure modes remain reliable — for now. This guide covers the tests that still work and the identity verification approach that is immune to any video manipulation.

What Deepfakes Can and Cannot Do in 2026

What they handle well

Blinking, nodding, smiling on request
Simple head turns left and right
Speaking while maintaining face coherence
Lighting changes that were in the training distribution
Standard background environments

What still breaks them

Hand covering or touching the face
Compound unscripted physical tasks named in the moment
Extreme lighting changes (torch shone at the face)
Moving to an unprepared background in real time
Holding a specific object that interacts with the face plane

The 5-Step Liveness Protocol That Still Works

Request the call with under one hour notice. Deepfake video injection requires setup time. An unscheduled call that the scammer cannot prepare for increases the probability of a refusal or a technical excuse.
Ask them to move to a different room immediately after the call connects. An unprepared background environment exposes deepfake boundary artefacts — the edges where the synthetic face meets the real background.
Name a specific object and ask them to hold it to their cheek. Choose the object during the call, not before. Watch the area where hand meets face for pixelation, blurring, or misalignment.
Type a sentence during the call and ask them to read it aloud. Watch lip movements at the syllable level. Deepfake lip-sync is accurate for common phoneme patterns but produces micro-errors on uncommon combinations.
Shine a torch or lamp directly at their face and ask them to move closer to it. Dramatic, sudden lighting changes that the model was not trained on produce visible degradation in real-time deepfake rendering.

Verify Identity — No Video Needed