Artificially created or altered audio, video or images that convincingly imitate real people, often using deep neural networks.
Also known as: Synthetic media, face-swap, audio cloning
In plain English
A deepfake is a fake video, image or audio made to look and sound real. Think of it like a very convincing Photoshop or face filter that can put someone else's face or voice into footage. People make them for fun, but they can also be used to mislead or deceive.
In practice
At work, deepfakes can appear as falsified interviews, manipulated evidence or misleading marketing that damage reputation and trust. When buying AI tools, teams should evaluate provenance, watermarking, detection capabilities and content policies to manage legal, compliance and brand risks.
Under the hood
Deepfakes are produced by generative deep neural networks—historically GANs and autoencoders and increasingly diffusion models—that learn mappings between identities, expressions and audio from large datasets. Training uses adversarial objectives, perceptual and temporal-consistency losses, and conditioning or fine-tuning to control identity, lip-sync and motion; post-processing and compression robustness are key engineering challenges.
Example
"We need to verify that the interview footage isn't a deepfake before publishing."