What are deepfakes, and how do you spot them?

The word combines "deep learning," the branch of AI behind today's image and voice generators, with "fake." A few years ago, convincing fakes took skill and hours of work. Now a short sample of someone's voice or a handful of photos can be enough to produce a clip that fools friends and family, which is why 2024, with elections in many countries, has people worried.

The clearest example so far this year arrived by phone. On January 21, two days before New Hampshire's presidential primary, voters received a recorded call in a voice that sounded like President Biden. It used one of his familiar expressions, then urged Democrats to skip the primary and save their vote for November. The advice was false: voting in a primary doesn't stop anyone from voting in the general election. The state attorney general called it an apparent illegal attempt to suppress votes, and the White House confirmed the president never recorded it. One 73-year-old voter told the Associated Press that the voice was so convincing she didn't question it at first. What tipped her off was the message itself, which didn't make sense.

Today, February 8, the Federal Communications Commission ruled unanimously that AI-generated voices count as "artificial" voices under the Telephone Consumer Protection Act, the main US law on unwanted robocalls. The ruling takes effect immediately and gives state attorneys general a new legal tool against voice-cloning calls.

Why your eyes and ears aren't enough

Researchers at the MIT Media Lab built a site called Detect Fakes so people could practise telling real videos from manipulated ones. Their main lesson is humbling: there is no single giveaway. A good fake can pass a quick glance, and tools keep improving. Physical clues still help, but the strongest checks are about where something came from and whether it makes sense.

Think of it as two layers. First, check the story around the clip. Then, if it still matters, check the clip itself.

A worked checklist

Here is how you might work through a suspicious item, using the New Hampshire call as the example.

For video, the MIT team suggests focusing on the face, since high-end fakes almost always alter it. Their questions, in short:

• Skin: does it look too smooth or too wrinkled, and does its age match the hair and eyes?
• Light: are shadows around the eyes and brows where you'd expect? Does glare on glasses shift naturally as the head moves?
• Hair and marks: do beards, moustaches and moles look real?
• Movement: does the person blink too little or too much? Do the lips match the words?

For still images, look at the same physical logic: light that comes from two directions, hands and ears that don't quite work, text in the background that turns into nonsense. For audio, listen for flat emotion, odd breathing or a voice that never stumbles, but treat those as hints only. A cloned voice on a phone line can be very hard to judge by ear, which is why steps 1 to 4 come first.

What should you do if one targets you?

If the fake uses your face or voice, or impersonates someone close to you, act on two fronts: protect people around you and build a record.

Start by warning the people most likely to be fooled. A quick message to family, colleagues or followers ("A fake video of me is circulating; I didn't say this") takes the power out of it. Families can agree on a check, such as calling back on a known number or using a private code word, before anyone sends money in response to an urgent call from a "relative." Scammers often clone the voices of family members precisely because panic makes people skip that step.

Keep evidence before it disappears: save the file or link, take screenshots showing the account that posted it and the date, and note the phone number if it was a call. Then report it. Social platforms have reporting forms for impersonation and manipulated media. Calls like the New Hampshire one can go to your state's attorney general; that office asked residents who received the call to report it to its election law unit. In the United States, the FCC's ruling means a robocall using an AI voice without consent now breaks federal law, and the same law lets individuals take robocallers to court.

Finally, keep your own response calm and factual. A short correction, posted where your audience will see it, usually does more than a long argument with the fake's author. If the fake is defamatory or used to extort you, a lawyer or the police are the right next step.