What is a deepfake? How fake videos are made and spotted

A deepfake is a video, image or audio recording that has been made or altered with AI so that a real person appears to say or do something they never did. The word joins "deep learning", the kind of AI that makes them, with "fake". Deepfakes can be harmless, as in film effects and satire, but they are also used for fraud, harassment and political deception, and they are getting harder to spot by eye.

In short

• A deepfake is built by training an AI model on real footage or recordings of a person until it can produce new ones in their likeness.
• The biggest harms so far are fraud, often by cloned voice, and sexual images made without consent; election fakes draw the most attention.
• Spotting glitches by eye is unreliable. Checking the source, and confirming through a channel you already trust, works better.

Where does the word come from?

The term spread in late 2017, when a user on the forum Reddit calling themselves "deepfakes" posted videos that pasted celebrities' faces onto other people's bodies, and shared the code. The technique wasn't new in research labs, but the free software made it something anyone with a good graphics card could do. Since then, the name has stretched to cover any realistic synthetic media of a real person, including cloned voices and wholly generated images.

How are deepfakes made?

Most deepfakes rest on neural networks trained on many examples of a person. The classic face swap works like this:

1. Collect hundreds or thousands of images of two faces: the person to be faked (the target) and the person in the original video (the source). Interviews, speeches and social media posts supply plenty.
2. Train a network to compress each face into a short numerical summary of its pose, expression and lighting, and to rebuild the face from that summary. The compression part is shared between the two faces; each face gets its own rebuilding part.
3. Feed in a frame of the source video, compress it, and rebuild it with the target's rebuilding part. Out comes the target's face, with the source's expression and head angle.
4. Repeat for every frame, then blend the edges and match the color so the new face sits in the shot.

Newer tools need far less. Generative AI models trained on millions of people can produce a convincing likeness from a few photos, and the same is true of voices. The US Federal Trade Commission warned in March 2023 that a scammer needs only "a short audio clip" of someone's voice, which may come from a video posted online, to clone it.

Kind	What changes	What it needs	Typical misuse
Face swap	One face replaced with another	Many images of both faces	Sexual images without consent, hoaxes
Lip sync	Mouth movements matched to new audio	Video of the person, plus new audio	Fake speeches and endorsements
Voice clone	A voice says new words	A short recording	Phone fraud, fake robocalls
Fully generated	A whole image or video from a text prompt	A general model	Fake news photos, impersonation

What harm do deepfakes do?

Fraud is the most costly. In early 2024, a finance worker at the engineering firm Arup in Hong Kong joined a video call with people who looked and sounded like the company's chief financial officer and colleagues. All of them were deepfakes. He went on to send HK$200 million, about US$25.6 million, across 15 transactions, according to police figures reported by CNN. Smaller versions are common: a cloned voice of a relative calling in a panic and asking for money, the twist on phishing the FTC warned about.

Sexual images made without consent are the most common deepfakes of all. In a 2023 count of 95,820 deepfake videos online by the security firm Security Hero, 98% were pornographic and 99% of the people targeted in them were women, as FactCheckNI reported in 2026. Elections are the third concern. In January 2024, Democratic voters in New Hampshire received robocalls with a voice imitating President Joe Biden, which the state attorney general's office said appeared to be artificially generated, telling them not to vote in the primary. The next month, the Federal Communications Commission ruled that calls made with AI-generated voices count as "artificial" under the Telephone Consumer Protection Act, which made such robocalls illegal without consent. Laws on deepfakes differ by place and change quickly.

There is also a quieter harm. Once people know any recording might be fake, a real one can be dismissed as fake too. Researchers call this the "liar's dividend": the existence of deepfakes helps anyone who wants to deny a genuine recording.

Are there legitimate uses?

Yes, and they use the same tools. Films use face and voice synthesis to de-age actors or finish a scene after an actor's death, with the estate's permission. Dubbing tools match an actor's lips to another language. People who are losing their voice to illness can bank recordings and have a synthetic voice made from them. Satire and art use obvious fakes to make a point. What separates these from abuse is consent and disclosure, not the technology.

How can you tell if something is a deepfake?

Looking for glitches helps less every year. Older fakes gave themselves away with odd blinking, blurred teeth, mismatched earrings or edges that shimmered when the head turned. Newer ones often don't. Detection software struggles too: in Meta's 2020 Deepfake Detection Challenge, 2,114 participants entered 35,109 models, and the best scored 82.56% accuracy on the public test videos but only 65.18% on a hidden set of new ones. A detector trained on yesterday's fakes misses tomorrow's.

What works better is checking the claim, not the pixels, the habits of critical thinking applied to media:

1. Ask where it came from. Find the earliest post, and the account or outlet behind it. A clip with no clear origin deserves doubt.
2. Look for confirmation. A real event of any size is usually reported by several independent outlets or filmed from more than one angle.
3. Check for content credentials. Some cameras, editing apps and AI tools now attach tamper-evident records of how a file was made, under a standard called C2PA, which some platforms display.
4. Confirm through a channel you already trust. If a boss, relative or bank asks for money or secrets by call or video, hang up and call back on a number you already have. Families and teams can agree on a code word in advance.
5. Notice pressure. Urgency, secrecy and strong emotion are the tools of scams whether or not a deepfake is involved.

Still argued

How much deepfakes have changed elections is debated. Some researchers and officials warn they are a growing threat to voters' trust; others argue that cheap edits and ordinary false claims still do more damage than sophisticated fakes, and that alarm about deepfakes can itself erode trust in real media. There is also disagreement on remedies: labeling and provenance standards, platform takedowns, and new laws each have supporters and critics, and none works alone.

Deepfakes are one part of a larger flood of synthetic media, along with the low-effort output called AI slop, and they are a reminder that artificial intelligence tools serve whoever uses them.

Check yourself

• In your own words, how does a face-swap deepfake turn one person's expressions into another person's face?
• Why does detection software tend to fall behind, and what should you check instead?
• What is the liar's dividend, and why does it matter even if a recording is real?

Sources

• British engineering giant Arup revealed as $25 million deepfake scam victim — CNN, 2024
• FCC makes AI-generated voices in robocalls illegal — Federal Communications Commission, 8 February 2024
• Scammers use AI to enhance their family emergency schemes — Federal Trade Commission, 2023
• Deepfake Detection Challenge results: An open initiative to advance AI — Meta AI, 2020
• Fake Joe Biden robocall in New Hampshire tells Democrats not to vote in the primary election — PolitiFact, 2024
• Content Credentials: verify media authenticity — Content Credentials, C2PA