Deepfakes & AI-generated media

How it works

A deepfake is media made or altered by machine learning: a face swapped onto someone else’s body, a voice cloned from thirty seconds of audio, a photo of an event that never took place. A few years ago this took a research lab. Today it takes a phone app, a subscription and an afternoon, and the results are good enough that “it looks real” has stopped being an argument.

The first widely noticed political deepfake came on 16 March 2022: a video of President Zelensky telling Ukrainian soldiers to lay down their arms appeared on a hacked Ukrainian news site and spread on social media. It was crude (stiff head, wrong voice) and was debunked within hours, partly because Zelensky immediately posted a real video. It still mattered, because it showed the playbook: produce the fake, push it through a trusted channel, hope it lands before the correction does.

Since then the technology has moved from politics into your pocket. Europol warned in 2022 that deepfakes were becoming a standard tool for fraud. In early 2024 an employee of the engineering firm Arup transferred about 25 million dollars after a video call in which the “chief financial officer” and several “colleagues” were all AI-generated. Cloned voices now phone grandparents asking for bail money. And in Hungary, AI versions of well-known TV presenters have spent the past few years recommending “investment platforms” in ads (see deepfake celebrity investment scams).

The point to remember: the tells are temporary. Six fingers, melting text, lips out of sync: every one of these has already been fixed in some tool. Pixel-hunting buys you a year at best. Provenance, where a thing came from, does not expire.

Example

@nation_first_daily, 1h [video: the opposition leader at a lectern, party logo behind him, 14 seconds] “The day we take office, the border fence comes down and the pension age goes up.” Leaked from an internal meeting, you will NOT see this on TV.

@marci_verifies, 40m The lectern and the logo are from his 2023 congress speech; the audio is new. No outlet has it, the party has not even been asked for comment, and “leaked” is doing all the work. Who filmed an internal meeting in 4K from the front row?

Notice the structure: a shocking sentence, a reason you cannot verify it (“leaked”, “not on TV”), and no origin. The reply does not inspect pixels at all. It asks where the clip came from.

Why it works on you

• Faces and voices are the oldest trust signals you have. Your brain spent a few hundred thousand years learning that a familiar voice means a familiar person. It has not had time to learn about cloning. It is the same reflex that powers authority bias.
• It arrives at speed, with a feeling. Fear (“the fence comes down”), urgency (“your son is in trouble”) or greed (“this platform made me rich”) are chosen precisely because they shut down checking. This is emotional hijacking in its purest form.
• It poisons the real thing too. Once everyone knows fakes exist, anyone caught on a genuine recording can say “deepfake”. Both effects, believing fakes and doubting evidence, serve whoever benefits from confusion.

What to do

When someone sends it to you: do not stare at the hands. Ask where it came from. Who posted it first, is it on the person’s or the outlet’s own channel, has anyone with a reputation to lose confirmed it? Run the SIFT method: stop, investigate the source, find better coverage, trace to the original. A real resignation, surrender or policy bomb is reported everywhere within minutes. A clip that exists only on one account is a clip, not news.

When a voice asks for money: hang up and call the person back on the number you already have. Agree a family code word in advance; it costs nothing. No bank, police force or ministry will ever ask you to move money during a call.

In a comment section: avoid “obviously fake”. It is not obvious, and people who were fooled will dig in. Say what you checked: “not on any official channel, no outlet has it, no source given.” Link the method, not the insult.

Don’t confuse it with

• Cheapfakes: slowed, cropped or re-dubbed real footage. No AI, often more convincing, far more common.
• Labelled AI art and parody. A clearly marked AI image of a politician riding a dragon is a joke. It becomes a deepfake when the label disappears and the caption says “real”.