What Is a Deepfake and How Can You Spot One?

Short answer: a deepfake is a video, image or audio clip that AI has generated or altered to make someone appear to say or do something they never did. Spotting one by eye is getting harder, so the best defence is a verification habit: check the source, look for the original, and confirm through a second channel before you trust or share.

What a Deepfake Is

The word joins “deep learning” and “fake”. A deepfake can be a face placed onto another person's video, a fully synthetic image of a person who does not exist, or a cloned voice. Not every use is harmful: film studios and accessibility tools use similar technology openly. The problem begins when it is used to deceive.

How They Are Made, at a High Level

AI models learn patterns from large amounts of real footage, photos or recordings of a person, then generate new content that matches those patterns. As the models and tools improve, results become more realistic and need less source material. That is why a few seconds of someone's public voice can now be enough for a convincing imitation. This post deliberately does not explain how to make them.

Why They Cause Harm

• Scams and voice-clone fraud: a call that sounds like a relative or boss asking for urgent money.

• Misinformation: fake clips of public figures spread before they are checked.

• Harassment: fake images or videos made of a real person without consent, which can be deeply damaging.

• Erosion of trust: when anything can be faked, real evidence can also be dismissed as “just a deepfake”.

Signs to Look For

These clues sometimes help, but treat them as hints, not proof. Newer fakes often show none of them.

• Odd blinking, stiff expressions, or lips that do not quite match the words.

• Blurry or shifting edges around the face, hair or ears.

• Strange lighting or shadows that do not match the scene.

• Hands, teeth, jewellery or background text that look distorted.

• A flat, oddly paced voice, or no natural breathing and background noise.

• The context: an extreme claim, a request for urgency or secrecy, or no trustworthy source.

Verification Habits That Work

1. Check the source. Who first posted it? Do reliable outlets report the same thing?

2. Search for the original. Use a reverse image search, or search key words from the clip, to find earlier or longer versions.

3. Call back on a known number. If a voice asks for money or codes, hang up and call the person on the number you already have.

4. Agree a family code word that only your close circle knows, to confirm identity in an emergency call.

5. Pause before sharing. Strong emotion is a reason to slow down.

What to Do If You Are Targeted

• Save evidence: screenshots, links, usernames and dates. Do not delete anything yet.

• Report the content to the platform where it appears and use its reporting tools.

• Tell people you trust, and your workplace or school if relevant.

• For fraud, contact your bank or service provider straight away.

• For harassment or threats, consider contacting local authorities or a support organisation.

What About Deepfake Detectors?

Detector tools analyse media for traces of AI generation and can be a useful extra signal. But they are imperfect: they can miss new kinds of fakes, flag real media as fake, and give a confident-looking score that is still wrong. Use them alongside source checking, never instead of it.

Common Mistakes

• Trusting a video because it “looks real”.

• Treating one detector's score as the final verdict.

• Sharing first and checking later.

Key Takeaway

You cannot rely on your eyes alone any more. Check where media comes from, find the original, and confirm urgent requests through a second channel.

For a deeper look at telling real from synthetic, read the chapter Deepfakes, Provenance and Disclosure. Also see what AI hallucination is for another way AI output can mislead.

Ready to start? Learn to use and judge AI responsibly, including how to tell real from synthetic. Begin with the free course AI Ethics and Responsible Use: Beginner.