A video appears online showing a public figure saying something outrageous. Or a voice message from a relative asks for urgent money. Either could be a deepfake: synthetic media created or altered by AI to imitate a real person. As the technology improves, knowing how it works and how to question what you see and hear is a basic digital skill.
What is a deepfake?
The word combines "deep learning" and "fake". A deepfake is audio, video or an image in which a person's face, voice or body is replaced or generated by AI to appear to say or do something they did not. Harmless examples exist, such as film effects, satire clearly labelled as such and consensual creative work. Harmful uses include fraud, harassment, non-consensual intimate imagery and political disinformation.
How deepfakes are made
Face swapping
A model learns the appearance of two faces from many photos or video frames. An autoencoder compresses faces into a compact representation and a decoder reconstructs them. By encoding one person's face and decoding with another person's decoder, the system produces the second person's face with the first person's expressions and head movements.
Face reenactment
The target's face is animated to match the expressions, lip movement and head pose of a driver video or audio. A single photo can sometimes be made to speak.
Fully generated images and video
Diffusion and other generative models can create realistic faces and scenes from text prompts, including people who do not exist.
Voice cloning
Text-to-speech models can reproduce someone's voice from a short sample, speaking arbitrary text with similar tone and rhythm. Quality has become high enough to fool family members in some reported scams.
Lip-sync and dubbing
AI changes mouth movements to match new audio, used legitimately for translation and misused to make people appear to say new words.
Why deepfakes are a problem
- Fraud: impersonating executives or relatives to request payments or sensitive information.
- Harassment and abuse: non-consensual explicit deepfakes cause severe harm, disproportionately to women.
- Disinformation: fabricated statements or events can influence opinion, markets or elections.
- Reputation damage: false clips can spread faster than corrections.
- The "liar's dividend": real evidence can be dismissed as fake, weakening trust in all media.
Signs that media may be synthetic
No single sign is conclusive, and high-quality fakes may show none. Look for combinations:
Visual clues
- Unnatural blinking or lack of it
- Mismatched lighting or shadows on the face compared with the scene
- Blurry or flickering edges around hair, ears and jawline
- Teeth that look blended or odd
- Skin that appears too smooth or inconsistent
- Jewellery, glasses or hands with distortions
- Backgrounds with warped lines or objects
- Lip movements not matching sounds exactly
Audio clues
- Robotic or flat intonation, odd pauses or breath patterns
- Background noise that cuts out unnaturally
- Mispronounced names or unusual phrasing for the person
- Audio quality that differs between segments
Context clues
- The content is shocking and designed to provoke strong emotion
- It comes from an unverified account with no reliable sources
- No other reputable outlet reports it
- The person's known behaviour or schedule does not fit
- The request is urgent and asks for secrecy or money
How to verify
- Pause before sharing. Emotional reactions fuel spread.
- Check the source. Who posted it first? Is the account credible?
- Search for corroboration from trusted news organisations or official channels.
- Use reverse image search on key frames to find earlier versions or originals.
- Look for the full clip and context, since short excerpts can mislead.
- Contact the person directly through a known number when it is a personal request.
- Use verification services and fact-checkers that investigate viral media.
- Check metadata and provenance information where available. Some cameras and platforms support content credentials (such as the C2PA standard) that record how media was created and edited, though absence of credentials does not prove a fake.
Detection technology
Researchers and companies build detectors that look for artefacts left by generators, inconsistencies in physiological signals or traces in frequency patterns. These tools are helpful but imperfect: they can miss new generation methods and produce false alarms, and it becomes an arms race. Watermarking of AI-generated content is another approach, but watermarks can be removed or may not be applied by all tools. Combine tools with human judgement.
Protecting yourself and your family
- Agree a family code word for emergencies, to verify urgent requests for money.
- Call back on a known number if you receive a voice or video request from a "relative", boss or bank.
- Limit public posting of high-quality voice and face data where practical, and tighten social-media privacy.
- Use strong account security and multi-factor authentication.
- Be sceptical of urgency, secrecy and unusual payment methods.
- Teach children and older relatives about voice cloning and fake media.
If you are targeted
- Save evidence: screenshots, links, account names, dates.
- Report the content to the platform using its impersonation or synthetic-media tools.
- Report to the police or cybercrime authorities, particularly for fraud or sexual content. In Nepal, the police cyber bureau handles online crimes.
- Seek support from trusted people and organisations; such abuse is not your fault.
- Consider legal advice about takedown and remedies.
For organisations
- Train staff to verify payment and credential requests through separate channels.
- Set approval rules for large transfers that cannot be satisfied by a single call or message.
- Monitor for impersonation of your brand and leaders.
- Prepare a communications plan to respond quickly to fake content.
Legal and platform responses
Many platforms ban deceptive synthetic media, and some countries have introduced laws against non-consensual deepfake imagery, election-related deepfakes or fraud. Rules vary widely and keep developing. Creating and sharing deepfakes of real people without consent can be illegal or lead to account bans.
Frequently Asked Questions
Can I always spot a deepfake with my eyes?
No. High-quality fakes can fool most people. Rely on verification of source and context, not only visual inspection.
Are all deepfakes harmful?
No. Consensual, clearly labelled uses exist in entertainment, education and accessibility. The harm comes from deception and lack of consent.
Can AI detectors prove something is fake?
They give probabilities, not proof. Use them alongside other evidence.
How can I tell if a voice call is cloned?
Ask a question only the real person would know, end the call and call them back on their known number, and be wary of urgent money requests.
Conclusion
Deepfakes use AI to imitate faces and voices convincingly. Because visual clues are unreliable, build habits of checking sources, verifying through separate channels and slowing down before reacting or sharing.