What is deepfake in AI?
A deepfake is synthetic media, a video, photo or voice recording, created or edited with deep learning so that a real person appears to say or do something that never happened.
The word joins "deep learning" and "fake". Old-style photo editing needed skill and hours of manual work. Deepfake tools learn a person's face or voice from examples and then produce new content automatically, often in minutes.
Common types you will see:
| Type | What gets faked | Typical misuse |
|---|---|---|
| Face swap video | One person's face placed on another's body | Fake celebrity ads, harassment |
| Lip-sync video | Mouth movements changed to match new words | Fake political or business statements |
| Voice clone | A person's voice copied from short audio | "Emergency" calls asking for money |
| Fully generated image | A realistic photo of an event that never happened | Viral misinformation |
Deepfakes are one part of generative AI, the same family of technology that writes text and creates images. The difference lies in intent: a deepfake is built to make you believe something false about a real person.
How are deepfakes made?
Deepfakes are made by training AI models on many images, videos or voice samples of a person until the model can generate new, realistic versions of that face or voice.
At a high level, without any step-by-step recipe, the main methods are:
- Face swap with autoencoders: a neural network learns to compress and rebuild faces, then rebuilds person A's expressions using person B's face.
- GANs (generative adversarial networks): two networks compete, one creating fakes and one trying to catch them, so the fakes keep improving. The idea comes from a 2014 research paper by Ian Goodfellow and colleagues.
- Diffusion models: the same kind of model behind popular image generators can produce or edit realistic photos and, increasingly, video.
- Voice cloning: text-to-speech models copy a speaker's tone and accent from recorded audio, sometimes only a short clip.
The quality depends heavily on the training data. Public figures with hours of video online are easier to fake convincingly, but ordinary people who post many photos and voice notes are not fully safe either. Many legitimate AI tools add guardrails that block impersonation of real people, though not every tool does.
How to spot a deepfake video, photo or voice
You spot a deepfake by checking the content itself for visual or audio glitches and, more importantly, by checking the source and context before you believe or share it.
Use this practical checklist:
- Face edges: look for blurring or flicker where the face meets hair, ears or neck.
- Lip sync: check whether mouth shapes match the words, especially on sounds like "p", "b" and "m".
- Eyes and blinking: stare, odd blinking or eyes that do not follow the head naturally.
- Lighting and shadows: the face lit differently from the background.
- Small details: teeth, earrings, glasses and fingers that warp or change between frames.
- Voice: flat emotion, no breathing sounds, strange pauses or a slightly robotic tone.
- Source: who posted it first? Is it on the person's official account or a trusted news site?
- Reverse search: run a screenshot through Google Lens or another reverse image search to find the original.
- Urgency: pressure to act now, send money or share quickly is a classic warning sign.
- Second channel: if a "relative" calls asking for money, hang up and call them back on their saved number.
Newer deepfakes hide many visual glitches, so steps 7 to 10 matter more every year. Detection tools exist, built on computer vision models, but none catches every fake. I tell students to treat the checklist as a habit, not a test: the question is not "does this look fake?" but "can I confirm this is real?"
Deepfake examples in India: scams and risks
In India, deepfakes mainly show up as financial scams, misinformation and impersonation that damages a person's reputation.
- Fake investment and product ads: videos that appear to show well-known business leaders, actors or cricketers recommending a trading app or scheme.
- Voice clone calls: a caller sounding like a family member claims an accident or arrest and asks for urgent UPI transfers.
- Fake video calls: scammers use altered faces or pretend to be officials to scare people into paying.
- Misinformation: edited clips of politicians or public figures that spread fast on WhatsApp, especially around elections or sensitive events.
- Harassment and morphed images: fake explicit or embarrassing images used to bully or blackmail, often targeting women and students.
These risks hit harder where people trust forwarded videos and voice notes from family groups. Deepfakes can also carry AI bias problems, since detection tools may work better on some faces and accents than others. Treat any shocking clip of a famous person as unverified until a trusted source confirms it.
What to do if you are a victim of a deepfake in India
If you are targeted by a deepfake in India, save the evidence and report it right away on the National Cyber Crime Reporting Portal (cybercrime.gov.in) or by calling the cybercrime helpline 1930.
Follow these steps:
- Do not pay or negotiate if someone is threatening or blackmailing you.
- Save evidence: take screenshots, copy the post links, note usernames, phone numbers and dates. Do not delete chats.
- Report on the platform: use the report option on Instagram, YouTube, WhatsApp, X or wherever it was posted. Platforms in India have grievance officers who must act on complaints.
- File a complaint: on cybercrime.gov.in, or call 1930, especially quickly if money was lost, because fast reporting improves the chance of freezing the transfer.
- Visit the police: you can also go to your nearest police station or cyber cell.
- Tell someone you trust: a parent, teacher or friend. Being targeted is not your fault.
Laws and platform rules around AI-generated content in India are being updated, so check the latest guidance on the official portal.
Is deepfake technology always bad? Legitimate uses
No, the same technology behind deepfakes has legitimate uses when the person involved gives clear consent and viewers know the content is synthetic.
- Films and dubbing: de-ageing actors or matching lip movements to a dubbed language.
- Education: historical figures "explaining" topics in a clearly labelled lesson.
- Accessibility: people who lose their voice to illness can use a cloned version of their own voice.
- Training and research: building better deepfake detectors needs examples of fakes.
The line between creative and harmful use comes down to consent, labelling and intent. If you want to understand the wider set of AI risks and safeguards, read about AI hallucination and explore the full AI glossary.