Glossary
What is a deepfake?
In short
A deepfake is a fake video, photo or voice recording made by a neural network so that a person seems to say or do something that never happened.
Images, video, audio · deepfake
MMOOZZGG.com editorsChecked: 2 min read
For a child
Sometimes you see a video online where a person you know says strange things, though in fact they never said them. A computer can fake faces and voices very well. So do not believe every clip or voice message right away, especially if it asks you to do something or send something to someone. Ask your parents first.
In plain words
Neural networks have learned to change a face in a video, lip movements and a voice in a believable way. Because of this, a photo, a clip or a call no longer proves that a person really said something. Fraudsters use such fakes: they call in the voice of a relative or a boss and ask for an urgent money transfer. A fake is hard to spot by eye, though unnatural movements, strange hands and teeth, wrong shadows or delays in conversation can give it away. A safer way: hang up and call the person back on a number you know yourself, or agree on a secret word with your family.
For business
For a company, fakes are dangerous in three places: a call or a video meeting in the name of a manager asking to pay an invoice urgently; a fake clip of the chief executive that hurts the company’s reputation; and tricking the remote check of a customer’s identity. In December 2024, the FBI warned that criminals pose as company executives in video calls using generated video. Protection rests on processes, not on staff instinct: a payment requested by voice is confirmed through a second channel, and large transfers need two signatures. Customer checks need a liveness check, not just a selfie. These rules belong in the AI risk policy.
For a specialist
A deepfake is synthetic or neural-network-altered audiovisual content that shows a real person. The classes are face swap, face reenactment, lip-sync to someone else’s speech, fully synthetic faces and voice cloning. Detection looks for artifacts at blending borders, inconsistent lighting, frequency traces of generation and audio that does not match the lips. The FaceForensics++ benchmark (Rössler et al., 2019), with more than 1.8 million altered images, showed that trained detectors clearly outperform people. Detectors transfer poorly to new generators, so the second line of defense is content provenance: the C2PA specification (version 2.1, September 2024) describes a manifest with file hashes and edit history, signed with a digital signature. Organizational measures are confirming requests through an independent channel, liveness detection and staff training.
Example
An accountant gets a call from a number that looks like the director’s, and a familiar voice asks her to pay a new supplier’s invoice urgently. Following the rules, she hangs up, calls the director back on a known number and learns that he asked no one for anything.
Related terms
Questions and answers
Sources
Not sure whether your company needs this?
AI solutions map