What Is a Deepfake in AI? How It Works, and How to Spot One

A deepfake is synthetic media created by AI models that swap a face, clone a voice or generate a person who never said the thing you are watching. This article explains the technology behind it in plain language, the Indian fraud cases that made it a real problem, how to spot one, and what the law says.

The plain definition

A deepfake is media, usually video, audio or a photo, where an AI model has replaced or synthesised a real person so convincingly that it looks like they said or did something they never did. The word comes from deep learning plus fake. That is the whole etymology.

The key part of the definition is the target being a real, identifiable person. An AI-generated image of an imaginary woman is synthetic media, not a deepfake. A video of the Prime Minister endorsing a trading app he has never heard of is a deepfake, and in January 2024 exactly that kind of clip circulated widely enough that the government issued an advisory about it.

Three flavours cover almost everything you will encounter. Face swap, where one person's face is mapped onto another person's body in existing footage. Lip sync, where a real video is kept but the mouth and audio are rewritten to say new words. And voice cloning, where a few seconds of someone's speech is enough to generate them saying anything, which is the variety currently costing Indian families actual money.

How the technology actually works

For years the standard method was a GAN, a generative adversarial network. Two neural networks are trained against each other. One generates fake frames, the other tries to catch them as fake, and the generator keeps adjusting until the detector can no longer tell. Think of a forger and an art inspector locked in a room for a million rounds. The forger gets very good.

Modern tools have largely moved to diffusion models and transformer-based audio models, the same family that powers image generators you have already used. They start from noise and denoise step by step towards a target, guided by whatever reference material they were given. The practical difference is enormous: a 2018 face swap needed hundreds of photos of the target and a decent GPU running overnight. A 2025 voice clone needs about fifteen seconds of audio from an Instagram reel and a free web tool.

That collapse in cost is the actual story. The technique is not new. The barrier to entry vanished.

Where deepfakes are causing real harm right now

Ignore the film-star mashups for a moment. The damage is concentrated in four places, and only one of them is what most people picture when they hear the word.

Financial fraud leads. In February 2024, a finance employee at a multinational in Hong Kong transferred about 25 million dollars after joining a video call where every other participant, including the CFO, was a deepfake. In India the same attack runs smaller and more often: a cloned voice calls a parent claiming to be their child in trouble, and the money moves before anyone thinks to call back.

Non-consensual intimate imagery is the largest category by volume, and the overwhelming majority of victims are women. This is not a side effect of the technology. For most of the last seven years it has been its main use.

Then political manipulation, especially audio clips released hours before a vote when there is no time to debunk them. And finally reputational attacks on ordinary people, which get no press coverage but wreck careers and marriages.

How to spot one

Detection advice ages badly, so treat the visual checks as a first pass, not proof. The blur-around-the-jaw and the never-blinking-eyes tells that worked in 2019 are mostly gone.

What still holds up reasonably well is listed below. But the single most reliable method is not technical at all: verify through a second channel. If your nephew calls asking for money, hang up and ring the number you already have saved. If a video shows a public figure saying something explosive, check whether any credible outlet has it. Real events leave more than one trace.

• Edges of the face, especially where hair meets forehead and where the jaw meets the neck. Swaps still smear here under movement.
• Teeth and tongue. Generators handle the inside of a mouth badly, and it often renders as a soft white block.
• Lighting mismatch. The face is lit from one direction, the shoulders and background from another.
• Audio texture. Cloned voices tend to have flat emotional range and no breath sounds, and background noise often cuts abruptly.
• Hands and jewellery in frame near the face. Fingers, earrings and spectacle arms still glitch.
• Ask yourself who benefits from you believing this, and why it reached you when it did.

What the law in India says, and what it does not

There is no standalone deepfake statute in India yet. What exists is a patchwork. Section 66D of the IT Act covers cheating by personation using a computer resource, with up to three years imprisonment. Sections 67 and 67A cover obscene and sexually explicit material. The Bharatiya Nyaya Sanhita provisions on forgery, defamation and cheating apply depending on the case. The IT Rules of 2021 require intermediaries to remove flagged impersonation content within 36 hours of a valid complaint.

In November 2023 MeitY issued an advisory to platforms tightening those obligations after a deepfake of actor Rashmika Mandanna went viral. In late 2025 draft amendments requiring visible labelling of AI-generated content were put out for consultation. The direction is clear enough: labelling obligations are coming.

Enforcement is the gap. Attribution is hard, the tools are hosted abroad, and most victims never file. If you are targeted, file at cybercrime.gov.in and report to the platform simultaneously, because the platform clock and the police clock run separately.

What this means if you work with AI, or want to

Understanding deepfakes is no longer a media-literacy nicety. If you are in finance, HR, journalism, law, banking operations or anywhere that verifies identity for a living, synthetic media has already changed the threat model of your job, and knowing what a model can and cannot fake is now part of doing that job competently.

The deeper point is that the same generative models behind deepfakes are the ones behind every legitimate AI tool you are being asked to use at work. Video generation, voice synthesis and image models are not separate technologies from the ones writing your reports. Once you have actually built something with them, the mystique drops away, and you get much better at judging what is plausible and what is fabricated. That hands-on layer is what most people are missing, and it is what the AI Masterclass sessions are built around for working professionals who need to use these tools rather than just read about them. Students and early-career folks who want a longer runway usually go the AI Fellowship route instead.

If you take one thing from this: stop treating video and audio as proof. For roughly a century, seeing was believing, and that century has ended. Verification now has to come from context and a second source, not from your eyes.

FAQs

1. Are deepfakes illegal in India?

There is no dedicated deepfake law in India as of now, but deepfakes are prosecutable under existing provisions, mainly Section 66D of the IT Act for impersonation, Sections 67 and 67A for obscene content, and the Bharatiya Nyaya Sanhita provisions on forgery, cheating and defamation. Platforms are also required under the IT Rules to remove flagged impersonation content within 36 hours.

2. What is the difference between a deepfake and AI-generated content?

All deepfakes are AI-generated content, but not all AI-generated content is a deepfake. The distinguishing factor is that a deepfake depicts a real, identifiable person doing or saying something they did not do, whereas a generated image of a fictional person or scene is simply synthetic media.

3. How much audio does someone need to clone my voice?

Current commercial and open-source voice cloning tools can produce a usable clone from roughly ten to thirty seconds of clear speech, which is less than one Instagram story. Longer samples improve accuracy, but they are not required for a clone convincing enough to fool a family member over a phone call.

4. Can deepfake detection software be trusted?

Detection tools are useful as a signal but not as a verdict, because detectors are trained on known generation methods and tend to fail on newer models they have never seen. For anything consequential, verification through an independent channel, such as calling the person back on a known number, is more reliable than any detector.

5. What should I do if a deepfake of me is circulating?

File a complaint at cybercrime.gov.in, report the content to the platform through its impersonation or non-consensual imagery flow, and preserve evidence with screenshots, URLs, timestamps and account handles before the post disappears. Doing the police complaint and the platform report at the same time matters, because the takedown timeline and the investigation run independently.

6. Do deepfakes always look fake if you watch closely enough?

No. High-effort deepfakes made with good source footage and manual cleanup can pass close visual inspection, and audio-only deepfakes are especially hard to catch because there are no visual artefacts at all. Assume that quality varies from obviously crude to genuinely indistinguishable.

Related reading

• What Is AI? Applications and Real Examples You Already Use
• what are the principles of responsible ai
• Is AI Really AI? An Honest Answer to the Question Everyone Is Asking
• Best AI Apps in 2025: What They Actually Do and Which Ones to Install
• AI Learning: What It Means, and How to Actually Learn AI in 2025