Deepfakes Explained: What They Are and How to Spot One

A deepfake is a video, image or audio clip that has been generated or altered by AI to show someone saying or doing something they never actually said or did. The term started with face-swap videos, but it now covers cloned voices, entirely synthetic photos and edited footage that looks convincingly real. The tools have become easy to use, which means anyone with a phone can run into one — in a news feed, a group chat or an unexpected phone call.

How deepfakes are made

Most deepfakes rely on machine learning models trained on large amounts of real images, video or audio. Given enough samples of a particular person, a model can learn to reproduce their face, expressions or voice and map them onto new material. Voice cloning in particular can work from fairly short recordings, which is why public videos, voicemails and social media clips can be enough raw material.

Common ways deepfakes are misused

• Scam calls — a cloned voice of a relative or boss asking for urgent money or a payment
• Misinformation — fake clips of public figures timed around elections, crises or breaking news
• Fake endorsements — celebrities appearing to promote investments or products they've never heard of
• Harassment — fabricated intimate or embarrassing images of real people

Visual clues worth checking

Older and cheaper deepfakes often give themselves away with small glitches. Look closely at edges where the face meets hair, ears or glasses, which may blur or shimmer. Lighting and shadows on the face may not match the rest of the scene. Teeth, hands and jewellery are often rendered oddly, and lip movements can drift slightly out of sync with the audio.

Audio clues

Checking still images

AI-generated photos have their own tells: backgrounds that melt into odd shapes, text on signs or clothing that turns into gibberish, mismatched earrings, or reflections that don't match the scene. Zooming in on these details is often more revealing than looking at the face. Some platforms and camera makers now attach content credentials that record how an image was made, which can help when they're present, though their absence proves nothing either way.

Cloned voices can sound flat, oddly paced or strangely clean, without the background noise you'd expect from a real call. Unusual pauses, a lack of natural breathing, or a voice that doesn't react normally when interrupted can all be signs something is off.

Why clues alone aren't enough

The honest caveat is that the best deepfakes no longer show obvious flaws, and the technology keeps improving. Relying on spotting a glitch is becoming less dependable over time. The stronger defence is checking context rather than pixels.

Habits that actually protect you

Ask where the content came from. A clip that appears only on an anonymous account, with no original source, deserves suspicion. If something dramatic is real, reputable news outlets or the person's official channels will usually confirm it. A reverse image search can reveal whether a photo has been lifted from somewhere else and edited.

For calls and messages, the rule is simple: if someone asks for money, passwords or secrecy under time pressure, stop and verify through a different channel. Call the person back on a number you already have. Some families and small businesses agree on a code word in advance so a quick question can settle any doubt.

Deepfakes work by exploiting trust and urgency. Slowing down, checking the source and confirming through a channel you control takes a minute — and it holds up even when the fake itself is flawless.

Frequently Asked Questions