deepfake detection

basicspreading

A collection of methods and tools used to determine whether a video, image, or audio recording was created or forged by artificial intelligence.

In plain terms

A deepfake is a video, image, or audio produced or altered by artificial intelligence to appear real: someone says or does something in it that they never did in reality. Deepfake detection is the antidote: it tries to determine whether a recording is original.

Imagine it like banknote verification. A cashier does not examine the entire note but looks at the watermark, the security thread, and the feel of the paper. Recognition tools similarly search for small telltale signs: blurred hair strands and facial edges, misplaced lights and shadows, audio lagging behind lip movements, or imperceptible pixel patterns. On today's high-quality deepfakes, these are often no longer visible to the naked eye.

There are two approaches. One examines the recording after the fact, looking for signs of forgery. The other verifies the origin: the camera or image-generating program adds a mark, a "digital birth certificate," to the file at creation time, which can be verified later.

Unlike traditional fact-checking, it does not examine whether a claim is true, but whether the recording itself is authentic.

How it works

Training: the recognition model is shown countless real and fake recordings so it learns the subtle difference between the two.
Artifact detection: the tool searches for telling traces in the examined recording, such as strange lighting, faulty face edges, unusual timbre, or pixel patterns characteristic of generative programs.
Provenance verification: it checks whether the file contains a watermark or credential data indicating what and when it was created and whether it has been modified since.
Probability specification: the result is not yes or no, but an estimate, for example, '85 percent chance of being AI-generated'.
Human judgment: in important cases, an expert also reviews the recording and compares it with the circumstances: who published it, where, and whether other sources confirm it.

An everyday example

You receive a video call in which your boss's face and voice request an urgent transfer to an unknown account. Your bank or company system can filter out the fake voice and image with deepfake detection, but the safest option is to hang up and call them back on the number you know.

Who is affected

everyoneOn social media and in messaging apps, anyone may encounter a fake video or audio recording impersonating a known person or even a family member.

bank customer, investorScammers use fake voices and faces to try to extract a transfer or pass video-based customer identification. → Money

entrepreneur, executiveFake calls and videos made in the names of company executives can cause serious financial and reputational damage.

lawyer, citizenIn election campaigns and in court evidence, it is increasingly common to question whether a recording is authentic. → Law

A common misconception

MythA deepfake detector program can definitely tell whether a video is fake.

In realityRecognizers only provide probabilities and make mistakes: they may let through a forgery or mark a genuine recording as fake. Their accuracy degrades particularly quickly against new generative tools.

Limits and risks

A cat-and-mouse game is underway: as soon as detectors learn a telltale sign, the next version of the generating program removes it.
Compression, resizing, or re-recording from the screen can erase traces and even the watermark, so the detection rate is lower on videos forwarded many times.
The absence of a watermark does not prove a recording is real, because not all generative tools label their output.
The existence of recognition also enables abuse: a real but embarrassing recording can be falsely labeled as a deepfake.

Similar, but not the same

fact-checking – Fact-checking examines the truth of statements, while deepfake detection determines whether the recording itself is original.
digital watermarking – Watermarking marks content in advance at the time of creation, while detection attempts to filter out fakes afterward, even without a mark.
AI text detector – An AI text detector tries to determine if written text was generated by a machine, while deepfake detection deals with images, video, and audio.

Related concepts

Broader

synthetic media, content authentication, Counter-disinformation measures

Narrower

voice clone detection, face swap detection, liveness detection

On the edge

deepfake, digital watermarking, voice cloning, generative AI, AI text detector

A brief history

The word 'deepfake' spread in 2017 when a Reddit user published face-swapping videos under this name. Research into detection accelerated afterwards: in 2019, the FaceForensics++ test database appeared, and in the same year, Facebook and its partners announced the Deepfake Detection Challenge competition.

Where things stand

By the end of 2026, banks, social platforms, and newsrooms are increasingly using detection tools. Article 50 of the EU AI Act, applicable from August 2026 (with transitional periods for certain systems), requires that AI-generated content be marked with machine-readable labels and that deepfakes be disclosed by the publisher. The regulation does not mandate the use of detection tools. As no reliable, universal detection tool exists, the focus is increasingly shifting toward provenance.

Status: 2026 November

In the news

Appears in 4 Agent Age stories; the curve shows weekly occurrence in the news magazine.

HKMA launches "ART" framework for responsible AI adoption in banking; banks test deepfake detectionOct 19, 2026 · Money
Ping An Digital Bank reaches 96% deepfake detection accuracy in HKMA GenAI sandboxOct 15, 2026 · Money
KPMG takes minority stake in deepfake detection firm Reality DefenderSep 10, 2026 · Money
Deel acquires deepfake-detection startup Clarity to counter fake hiresAug 3, 2026 · Money

Also known as

deepfake detection · deepfake detection · deepfake detection · deepfake filtering · deepfake detection · deepfake detection · deepfake detector · synthetic media detection

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.