AI 101 — What is a deepfake?
A deepfake is video, audio, or image content in which AI makes a real, identifiable person appear to say or do something they never did. The name is short for "deep learning" plus "fake," and the second half of the definition is the part that matters: a deepfake is not simply AI-generated media. It is a real person placed in a false situation — your face, your voice, a chief financial officer's on a video call.
Why it matters right now. The word has been doing heavy lifting in our coverage all month without much explanation. On Friday we covered Particle6's synthetic performer Tilly Norwood, whose entire product is a likeness of a person who does not exist. This month a Chinese court published guidance on how to litigate deepfake cases, Apple shipped a way to prove an iPhone photo is not one, and the EU's labelling rules for manipulated content took effect on 2 August 2026. All four stories assume you know what a deepfake is. This is the ten-minute version.
The term itself is only nine years old. An anonymous Reddit user going by "deepfakes" started posting face-swapped clips to a subreddit of the same name in November 2017, using open-source machine learning libraries. Within three months a free desktop app called FakeApp put the technique in anyone's hands, and Reddit, Twitter and Pornhub banned the resulting communities in February 2018. Everything since has been scale, not invention.
The mental model
Two families of model do the work. The classic face-swap pipeline is a pair of networks that share a narrow middle layer: one squeezes a face down into a small code, the other rebuilds a face from that code. Train the rebuild step on thousands of images of your target, and the same code now comes back out wearing their face — then a third step finds the face in every frame of the source video and pastes the new one in. Modern systems skip most of that and use diffusion models, which learn to generate an image by removing noise step by step. Give one a photo of a person plus a description — "standing in an office, mid-sentence, looking left" — and it paints them there.
Either way, the model never saw the event. It learned from examples what a face looks like at a given angle, in a given light, wearing a given expression. What it holds is a statistical impression of a person, not a copy of a recording. That distinction explains both what deepfakes are good at (likeness, in familiar framings) and where they break (behaviour the training data never showed — answering a question nobody has asked before).
The understudy analogy
Think of an understudy in a theatre who has watched the lead actor for a thousand performances. They have the voice, the walk, the raised eyebrow. Put them on stage mid-scene, in costume, under lights, and the audience cannot tell — for a few minutes, doing exactly what the script says. Now ask them to improvise a press conference. The impression falls apart instantly, because an impression is memorised surface, not a person. That is precisely the shape of the technology: convincing in short, scripted, well-lit moments, and hollow the second it has to think. It is also why the AI "actress" Tilly Norwood froze when a journalist asked her a question about her own film — the likeness held, the mind behind it did not.
Common misconceptions
"Deepfakes are video." Audio is the fastest-growing and cheapest kind. A voice clone needs seconds of sample audio, which is why the classic scam is now a phone call from a relative in trouble — a real voice, a fake emergency, a payment request.
"You can spot them." The 2018 tells — strange blinking, mangled hands — were artefacts of small models. A single still image or a ten-second audio clip is now essentially undetectable by eye, and detection degrades further every time a file is compressed, re-encoded, or re-recorded off a screen. The visible failures show up in unscripted conversation, not in the pixels.
"Detection tools will save us." Detectors work, sometimes, on watermarked or unedited output, and lose accuracy as the content is passed around. That is why the regulatory push has moved to provenance instead: machine-readable marks and signed metadata that travel with the file. It is stronger evidence than a guess, but it only covers content that a cooperating model produced — What is AI watermarking? goes through the technical side.
"It is a celebrity problem." The most expensive confirmed case is a finance worker at the engineering firm Arup who transferred $25 million in early 2024 after a video call with what turned out to be deepfaked colleagues, including the company's chief financial officer. The most common victims are private individuals: non-consensual intimate imagery is the single largest category of deepfake abuse, which is why the US TAKE IT DOWN Act now requires platforms to remove it within 48 hours of a valid request, with the FTC enforcing since May 2026.
"Any AI image is a deepfake." Not quite. A synthetic scene of nobody in particular is a different legal and ethical object from a fake recording of a specific named person. The EU's AI Act labels the second one explicitly and demands that deployers disclose it, with a lighter requirement for evidently artistic or satirical work.
Where to learn more
Start with the EU AI Act's transparency rules on Article 50 — the clearest published account of what must be labelled, by whom, and from when. Then look at the two enforcement mechanisms already live: the FTC's Take It Down reporting portal in the US, and China's labelling measures, in force since September 2025, which require both a visible label and metadata marks that platforms must check and that nobody may strip.
Related reading: What is AI watermarking? · AI 101 — What is AI regulation? · Tilly Norwood's press tour sold a chatbot, not an actor
Should a convincing fake of a real person be illegal to post by default, or should the burden stay on platforms to label and remove it? Tell us in the comments.
