What is a deepfake?

A deepfake is media in which a person's likeness — face, voice, or full body — is synthetically generated or transferred without the person performing the action shown. The 'deep' refers to the deep neural networks that made the technology accessible starting in the late 2010s.

Face-swap deepfakes replace one person's face with another's in existing footage. Voice-clone deepfakes synthesize a target speaker's voice to say new words. Full-body synthesis generates a person from scratch — no source footage required.

Each type leaves different artifacts. Face-swaps break around the jawline and ears; voice clones flatten emotional contour; full-body synthesis produces uncanny motion and wardrobe inconsistency.

Detectors look for patterns the human eye misses: temporal warping, lip-sync drift, prosody flattening, and frequency-band quirks. They produce a probability, not a verdict. On well-known production deepfakes, modern detectors usually flag the manipulation. On adversarially-tuned deepfakes designed to evade detectors, performance drops sharply.

It cannot prove a video is real — only that the engine found no synthetic markers. It cannot identify which deepfake tool was used. It cannot evaluate whether the underlying claim is true; an unedited video presented out of context is often more harmful than any deepfake.

Treat detector output as one input. Combine it with provenance (where did the file come from?), context (does the claim match other reporting?), and behavior (is the alleged speaker known to make claims like this?). The combination is far stronger than any single signal.