What a deepfake is, and how to detect AI-generated fake video, audio and images
A deepfake is fake video, audio or imagery created or altered with artificial intelligence, making a real person appear to say something they never said or do something they never did. This explainer sets out how the technology works, real cases such as the 2024 US$25 million fraud, the practical signs for detecting synthetic content, and how to protect your family — on the basis of sources only.
A deepfake (from "deep learning" and "fake") is fake video, audio or imagery created or altered using artificial intelligence, in particular deep-learning algorithms. A real person's face, voice or movement is reworked so that they appear to say something they never said or do something they never did. The term emerged online in 2017. Only sourced facts are set out below.
How real is the threat
In January 2024 a finance worker at the Hong Kong office of the British engineering firm Arup was defrauded over a video call in which every other participant — including the company's chief financial officer — was a deepfake. As a result, about US$25.6 million (200 million Hong Kong dollars) was transferred to the fraudsters through 15 transfers in a single day. Hong Kong police reported the case in February 2024, and Arup later confirmed it was the victim.
In 2023 an AI-generated fake image showing Pope Francis in a white puffer jacket went viral and fooled millions — a vivid example of how quickly AI-generated fake imagery can spread. Experts note that voice cloning — faking a real person's voice — is the fastest-growing category of AI fraud in 2026: fraudsters may fake a relative's voice and ask for money under the pretext of an "emergency."
How to check a video
According to experts' 2025–2026 guidance, fakery often shows at the boundaries — the hairline, the jaw, where the neck meets the collar — where there may be blurring, flickering or a poor join. Current models may render a face as over-smoothed, "waxy" and unnaturally clean. Lighting and shadows that don't match the surroundings, unnatural blinking, hand movements, and poor lip-sync on "p," "b," "t" and "d" sounds can also be signs.
How to check audio
AI-generated audio often has an even, robotic cadence: words at equal spacing, without natural signs such as breathing, hesitation or self-correction. On a suspicious call, asking the person to slowly turn their head to the side or move closer to the camera still trips up many current models — a flaw appears in the image.
The most important rule — verify the source
A genuine, significant event is always covered by several independent sources. If a video shows something world-shaking but no credible outlet confirms it, that is a strong sign of fabrication. Checking an image with a reverse image search, comparing official accounts, and pausing on urgent money requests to confirm through a second channel (for example, calling back a known number) are practical defences. Agreeing a secret "safe word" within the family also guards against fraud.
Detection tools and the law
In 2026, deepfake-detection tools combine AI classifiers, metadata forensics and physiological-signal analysis, but no single tool is 100% accurate — detection and generation are in a continuous "arms race." To confirm where content came from, the C2PA (Content Credentials) standard is being developed: it attaches cryptographic provenance data to an image or video. The European Union's Artificial Intelligence Act, adopted in 2024, provides for an obligation to label AI-generated or altered content (deepfakes).
In sum: deepfakes are a fast-developing, real threat already used in fraud. For now no tool can detect fakery with 100% certainty, so the most reliable defence is to watch for flaws in image and audio, verify the source across several independent outlets, and confirm urgent money requests through a second channel. This explainer gives the general facts.
