What is a deepfake, and how can anyone tell what's real?

A deepfake is an image, video or audio recording that has been edited or generated using artificial intelligence so that it shows or says something that never happened. The word blends deep learning and fake. The technique is not new: an early research system called Video Rewrite, published in 1997, modified footage of a person speaking so that they appeared to mouth words from a different audio track. A key technical turning point came in the mid-2010s with generative adversarial networks.

Deepfakes can be used for scams. In 2019, a UK-based energy firm's CEO was reportedly tricked over the phone into transferring €220,000 by someone who used audio deepfake technology to impersonate the voice of the parent company's chief executive.

Telling what is real is hard, and even detection software and people struggle. The winning model of a major detection challenge was 65% accurate on 4,000 held-out videos, and an MIT paper found that ordinary people were 69 to 72% accurate on a sample of 50 of them. Detection is also criticized as a moving goal post, because fakes improve as detectors do.

So a second approach has grown: instead of spotting fakes, label real content. A standard called C2PA attaches signed information about where a file came from, and AI watermarking changes the output of an AI model so it can later be identified. Neither proves that something is true, only where it came from or how it was made.