How Deepfakes Are Made, and 9 Ways to Spot a Fake Video or Voice

A video of a celebrity endorsing an investment scheme. A voice note from your boss urgently asking for a payment. A clip of a politician saying something shocking. Increasingly, these are deepfakes: AI-generated media designed to look and sound real. Knowing how they are made is the best defence.

How deepfakes are created

• Face swapping: an AI model learns a person's face from many photos and videos, then replaces another person's face in a video frame by frame.
• Lip-sync: the mouth in a real video is regenerated to match new audio.
• Voice cloning: a model learns a voice from a short recording and can then speak any text in that voice.
• Fully generated video: newer models create entire scenes from text descriptions.

Most of these tools are built on the same family of AI models used in image generators. Read how AI image generators work to understand the core idea.

Why they are getting harder to spot

Early deepfakes had blurry faces, strange blinking and robotic voices. Modern models fix many of these flaws, and compression on social media hides the remaining artefacts. Relying on your eyes alone is no longer enough.

9 ways to spot a fake

1. Check the source. Is it from the person's verified account or a reputable news outlet?
2. Watch the edges of the face, hairline and jaw for shimmering or blurring when the head turns.
3. Look at teeth, ears and hands, which models still sometimes render oddly.
4. Notice lighting mismatches between the face and the background.
5. Listen for flat emotion, odd pauses and missing breathing sounds in voices.
6. Check lip sync on sounds like "p", "b" and "m", where lips must close.
7. Search for the clip or key frames using a reverse image search.
8. Be suspicious of urgency. Scams push you to act before thinking.
9. Verify through a second channel. Call the person back on a number you already know.

Family safety rule: agree on a private code word with family members. If someone calls in panic asking for money, ask for the code word.

How platforms and engineers are fighting back

• Content credentials that attach a tamper-evident record of how an image or video was made.
• Invisible watermarks added by some AI tools to their outputs.
• Detection models trained to recognise synthetic media, though this is an ongoing race.

Frequently asked questions

Is making a deepfake illegal?

Using someone's likeness to defraud, defame or harass is illegal in many countries, and new rules specifically targeting harmful deepfakes are being introduced. Laws vary by country.

What should I do if I see a deepfake of myself?

Save evidence, report it to the platform, and report fraud or harassment to the police or your country's cybercrime portal.