The sophistication of AI-generated media has reached a point where a real-time video call with what appears to be a family member, a colleague, or a company executive may not involve that person at all. In 2026, deepfake video and voice cloning technology has become accessible enough that fraudsters are using it in targeted scams against ordinary people — and the financial losses are significant.
This guide gives you the practical knowledge to identify AI-generated media and protect yourself from the most common attack patterns.
How Modern Deepfakes Work
Understanding the technology helps you identify its weaknesses. Modern video deepfakes use generative adversarial networks and diffusion models to map one person’s facial expressions and movements onto another person’s face in real time. Audio deepfakes — voice clones — require as little as three seconds of a person’s real voice to generate a convincing replica that can say anything.
Both technologies have improved dramatically, but they still have tells — artifacts left by the generation process that are invisible to someone not looking for them but obvious once you know what to watch for.
Visual Tells in AI Video
Train your eyes to notice these specific patterns when on a video call or watching a video message:
- Unnatural blinking — Early deepfakes rarely blinked. Current models blink, but the timing is often slightly irregular or the blink duration is too uniform. Natural blinking is variable.
- Edge artifacts around the face — Look at the hairline and the boundary between the face and the neck. Deepfakes often produce a subtle halo, slight colour banding, or a soft blur that does not match the sharpness of the rest of the frame.
- Inconsistent lighting on the face — If the background lighting changes — a window, a moving light source — the face in a deepfake often fails to update its lighting consistently. The face may appear to have its own independent light source.
- Teeth and mouth interior — The inside of the mouth, teeth texture, and tongue movement are still among the hardest things for generative models to render convincingly. Look closely during speech.
- Unnatural head movement — Deepfakes tend to keep the head relatively still or move it in constrained arcs. Natural human head movement is more varied and spontaneous.
Audio Tells in Voice Clones
- Flat emotional range — Voice clones reproduce the tonal qualities of a voice accurately but often flatten the emotional dynamics. Stress, excitement, and sadness are harder to clone convincingly than neutral speech.
- Slightly unnatural pacing — Real speech has micro-pauses, filler words, and rhythm variations tied to thought. Cloned voices often sound slightly too fluent or have pauses in slightly wrong places.
- Background audio mismatch — A cloned voice call may have inconsistent background noise — a room tone that does not quite match the claimed location, or background noise that cuts in and out unnaturally.
Behavioural Red Flags
Technology aside, scammers using deepfakes follow predictable behavioural patterns:
- Creating urgency — “I need you to transfer money right now”, “Do not tell anyone about this yet”
- Requesting actions outside normal channels — a CEO asking for a wire transfer via WhatsApp rather than through the finance system
- Avoiding live interaction — refusing to turn on video, refusing to answer spontaneous questions, asking you to call back on a different number
Practical Protection Measures
Establish a family safe word — a word or phrase known only to close family members that must be used to verify identity before any urgent request involving money or sensitive information is acted on. Ask the caller to say the safe word. A deepfake cannot comply because it does not know what the word is.
For business contexts, implement a dual authorisation policy for any financial transfer regardless of who requests it. No single video call or voice message should be sufficient to authorise a payment.
Stay sceptical, stay slow, and remember: legitimate urgent requests can always wait 60 seconds for a verification call back on a known number.
