Deepfake

How is a deepfake made?

Early deepfakes (2017) used autoencoders trained on two faces to swap one onto the other frame by frame. Current tools use diffusion models and generative adversarial networks that need a single reference photo, and voice models that clone speech from seconds of audio. The result can be a swapped face in an existing video, a lip-synced clip, a cloned voice on a phone call, or a wholly new image of a real person in a scene that never happened. Live face swaps on video calls are the newest form and the one behind many 2026 fraud reports.

How is it different from a generated face?

A generated face that belongs to nobody (the "this person does not exist" kind) is synthetic but not a deepfake, because no real person is depicted. A deepfake depicts someone real. The distinction matters for law: most deepfake statutes protect identifiable people, while fully synthetic faces mostly raise fraud questions, as on dating profiles (see catfishing).

Which laws cover deepfakes?

How are deepfakes detected?

Three families of method: artefact detectors that look for inconsistencies in lighting, blinking, teeth, ears and the boundary of a swapped face; provenance signals such as C2PA content credentials that record how a file was made; and liveness detection on live calls and identity checks. None is reliable alone. Detector accuracy falls on compressed, resized or re-encoded media, which is what social platforms produce.

What can a face search do about a deepfake?

A reverse face search can find where a deepfake of a real face has been published on public pages, because the synthesised face still resembles the person. It cannot tell that an image is synthetic, cannot see private channels, and finds nothing for a face that was never real. For removal, the NCII removal directory and the report impersonation directory list the routes.