How do you tell if a video call is a deepfake in India?
Key takeaways
- The I4C warned in June 2026 that deepfake video defeats facial authentication, liveness detection and video-based KYC, so a liveness check is one factor, not two.
- A static replay fails a live challenge, so asking the caller to turn their head or move an object is a stronger test than any visual artefact.
- SEBI's caution of 17 July 2026 treated the video itself as unreliable, which is why the verification step has to leave the channel the request arrived on.
- A family safe word and a callback to a number already on file cost nothing and stop most voice and video impersonation attempts.
- For a business, dual authorisation on any new beneficiary and a 24-hour cooling-off period are the controls that actually prevented losses in the 2026 Indian cases.
As of 24 October 2026: the service descriptions, prices and turnaround figures on this page are current.
What can you test live in a video call?
Five things, and each one defeats a static replay, which is the most common form of the fraud. Ask the person to turn their head sharply to one side, which forces a generative model to synthesise the occluded region. Ask them to hold an object up to the camera, and to move it, which breaks the consistency of a fixed background. Ask them to pass a hand in front of their face, which tests depth and occlusion. Ask them to say a random word or read a short random phrase, which a lip-sync model handles poorly. And ask them to look off camera and back. Real humans are slightly late, slightly off-centre and slightly uneven; generated faces are often too perfect.
What visual artefacts should you look for in the frame?
Six, in rough order of reliability. Blink cadence that does not match the speech, or eyes that do not blink at all. Lip-sync drift on plosives and closed-lip words. Skin texture that is too smooth, with the micro-detail and specular pattern of real skin missing from the forehead and cheeks. A hard or smeared boundary at the jaw, hairline and collar, where the composite sits. Background warping, especially around reflections, glass, text on a wallboard and moving objects. And over-smoothed audio, where a cloned voice has unnaturally uniform background noise and no room reverb. None of these is a verdict, and modern generators fix several of them. Treat them as reasons to escalate to a callback, not as proof of fraud.
What questions should you ask to verify the person is real?
Three kinds, and they work because they are not in the video. A shared-memory question: something only the two of you know, and never discussed online. A context question with a wrong-premise trap, such as asking what the other person is doing tomorrow and checking the answer against reality. And a channel test: hang up and call back on a number you already have. If the caller answers on the second call, it was almost certainly the real person, because the fraudster is on the first call. For money, add the safe word. A good shared-memory question cannot be answered by a model trained on public data, and it cannot be answered by a deepfake at all, because the person generating the video does not have the memory.
What should you do if you suspect a video call is a deepfake?
Five steps, and speed matters. Do not disclose that you suspect it, because the caller will correct their behaviour. Stay on the line, or at least keep recording if your policy allows, and note the time, the number and the platform. Do not confirm any payment, any credential or any new beneficiary. Hang up and call back on a number already on file, and if it does not ring, you have your answer. Report the attempt to your bank's fraud helpline and, if money moved, to cybercrime.gov.in and helpline 1930 the same day, and to CERT-In within 6 hours if your organisation was targeted. And preserve the video, the screenshots and the call log, because the takedown and the case both depend on the original file.
| Check | How to run it | What it catches |
|---|---|---|
| Head turn | Ask the caller to turn sharply to one side | A static replay or flat generative face with no side profile |
| Object movement | Ask them to hold and move an object near the camera | Inconsistent rendering of a moving object and its shadow |
| Occlusion | Ask them to pass a hand in front of their face | Depth failure and background fill artefacts |
| Random phrase | Ask them to read a random short phrase aloud | Lip-sync drift that only appears on unscripted speech |
| Shared memory | Ask something only the two of you know | Any clone, because the generator does not have the memory |
| Callback | Hang up and call a number already on file | The whole fraud, because the second call goes to the real person |
What are the frequently asked questions?
Get a straight answer on this
Send us a recording or a call detail from a call you are unsure about, and we will run a forensic assessment of the media and the workflow around it.
Send us the details See pricing