What Is a Deepfake? How It Works and How to Spot One

What it actually is

The EU AI Act, the regulation governing artificial intelligence across the European Union, defines a deepfake in Article 3(60) as AI-generated or manipulated image, audio, or video content that resembles existing (or plausible) persons, objects, places, entities, or events and gives a false impression of their authenticity. Three elements have to coexist: resemblance to a real or plausible subject, the existence of that subject, and the content’s ability to look authentic to whoever sees or hears it. UNESCO places it within a broader category, synthetic content: any material, audio, image, or video, produced by generative artificial intelligence, of which the deepfake is the most deceptive case because it convincingly mimics a real person’s voice or likeness.

How it’s made, in brief

Until around 2022, the leading technique was generative adversarial networks, or GANs: two neural networks pitted against each other, a generator that produces a fake image and a discriminator that judges its authenticity. The process repeats for hours or days, until the generator produces results the discriminator can no longer catch. Since 2023, this architecture has largely given way to diffusion models, the family that includes Stable Diffusion, Sora, and Runway Gen-3.

In practice, a few seconds of real voice or a handful of images are enough to train a model that can generate a video within minutes: some commercial tools, such as HeyGen, make the process accessible to anyone with a starting video of the person being imitated. Within a few years, machine learning applied to this purpose went from a research lab to a widely available tool.

A real case: the video call that cost $25 million

Practical example: in January 2024, an employee at a Hong Kong engineering firm joined a video call with what looked like the chief financial officer and several other colleagues. They were all deepfakes. Convinced he was talking to real superiors, the employee authorized transfers totaling $25 million. Analysts at Trend Micro believe it’s unlikely the fraudsters used an AI that responded live; more probably, they played a set of pre-recorded clips in real time during the call.

Misuses beyond disinformation

The 2022 report from the Europol Innovation Lab lists a range of criminal uses for deepfake technology that goes well beyond false news: online harassment and humiliation, extortion and fraud like the case above, document forgery, digital identity theft, non-consensual explicit material, child sexual exploitation, manipulation of evidence in legal proceedings, and finally disinformation and political polarization. A Recap on spotting fake news covers a different problem: fake news is a false story in any form, while a deepfake is synthetic content built with AI to look real.

Why eyes and ears aren’t enough anymore

A study published in Nature Communications in 2024, led by Matt Groh and colleagues at the MIT Media Lab, tested how well people can recognize a political deepfake using a set of 32 short speeches by Presidents Joe Biden and Donald Trump. On the audio side the picture is even starker: a 2025 study cited by UNESCO found that people often perceive an AI-generated voice as identical to a real one, without being able to tell them apart reliably. UNESCO also points to a parallel limit on the technology side: automated detection tools consistently lag behind the techniques used to create deepfakes.

How to spot a deepfake, in practice

Some precautions remain useful precisely because they don’t depend on catching a visual or audio detail. UNESCO suggests that families agree on a code word to verify each other during a suspicious call, and proposes the “prove you’re live” technique: asking the other person to perform an unplanned physical action in real time, something live-generation systems struggle to reproduce naturally.

The rules: what the EU AI Act requires

Since August 2, 2026, the AI Act has required anyone who distributes a deepfake, the so-called “deployer” of the system, to disclose it clearly to whoever views or hears it, at the latest upon first contact with the content. Works that are evidently artistic, creative, satirical, or fictional are exempt from the strict version of the rule, needing only a disclosure that doesn’t get in the way of enjoying the content. Content generated before August 2, 2026, doesn’t need to be labeled retroactively, even though the European Commission recommends it. For systems already on the market, the technical marking obligations begin on December 2, 2026. Non-compliance carries penalties of up to €15 million or 3% of worldwide annual turnover. The rules are covered in more depth in the Recap on the EU AI Act.