What is a deepfake?

A video in your feed shows a famous actor recommending a skin cream. The face looks right. So does the voice. But the actor has never mentioned this product, and never made this video. Someone else did, with a few tools and a reason to borrow that actor's trust. That's a deepfake. The word joins deep learning, a kind of artificial intelligence (AI), with the word fake.

The Government Accountability Office (GAO), which researches technology for the United States Congress, describes a deepfake as a video, photo, or audio recording that seems real but has been manipulated with AI. The key is a real person, shown doing or saying something they didn't. A cartoon made with AI doesn't count. A fake clip of your mayor does.¹³

It starts with recordings of the person. A generative model, the kind of AI that makes new images and sound, is trained on pictures or clips until it can rebuild that face or voice. Then it can place the face on someone else's body, change the mouth to match new words, or speak in that voice. In 2024, the Federal Bureau of Investigation warned that these tools cut the time and effort criminals need to fool people.¹⁵

Most deepfakes fall into three kinds.

• Face swaps, where one person's face is placed on another person's body.
• Altered speech, where the mouth is changed to match words the person never said.
• Cloned voices, where only the sound is faked, often in a phone call.

Topic 15, on voice scams, covers that last kind.¹

The technology also has honest uses. The GAO lists entertainment among them, where a voice or likeness can be used in a film for a creative effect, such as making an actor look younger in a flashback. And voice banking helps people who are losing their speech to an illness keep a synthetic version of their own voice. The tool is the same. What changes is who uses it, and why.¹⁴

A deepfake doesn't make itself. A person decides whose face to borrow, what to make them say, and why. Most of the harm comes from those choices: a scam, a smear, a fake endorsement. That's why the rules are aimed at the people who make and share them.

In the European Union, the AI Act now requires that many deepfakes be disclosed as artificially made. The rule has applied since August 2026, with a lighter touch for work that's clearly artistic or satirical. Other places are passing labeling laws of their own. Even so, a label only helps when it's there, and plenty of clips arrive without one.³

You can't always tell by looking, and you don't need to become a detective. The better question is where a clip came from. Who posted it first? Does the person's own account, or a news outlet you trust, show the same thing? Next time a clip surprises you, take ten seconds to find its source before you share it.²