← AI Terminology

Deepfake

A deepfake is synthetic media (often faces/voices) generated or manipulated by deep learning to appear real, frequently for impersonation or entertainment.

The term covers face swaps, full synthetic video, and cloned audio.
Why It Matters in AI
Generative models enable both creative tools and fraud/disinfo risks. Detection, provenance (C2PA), and policy responses are now first-class AI safety and trust topics.
Key Points
Aspect Description
Eval Detection AUC; human fool rates
Risks Fraud, non-consensual porn, political disinfo
Methods Face-swap GANs/diffusion, voice cloning, lip sync
Related Voice cloning, generative models, watermarking
Defenses Detectors, watermarking, cryptographic provenance
Positive Film VFX, avatars with consent
Simple Analogy
A rubber mask so perfect that video calls can no longer prove who is speaking — powerful costume, dangerous in the wrong hands.
Common Usage Examples
  • Face-swap research history (autoencoders/GANs)
  • Audio deepfake fraud call centres
  • C2PA content credentials
  • Platform deepfake labeling policies
Summary
In short: Deepfakes are AI-generated or swapped media that mimic real people — a dual-use capability driving detection and provenance efforts.