Glossary

Definitions of terms used across the Deepfake Research Index. Each entry stands alone.

Term	Definition
Deepfake	Synthetic or manipulated audio, image or video, produced with deep learning, that depicts a person saying or doing something they did not. Many papers also use the term for wholly synthetic faces of people who do not exist.
Synthetic face / AI face	A face image generated from scratch by a model (for example StyleGAN2 or a diffusion model), depicting no real person.
Face swap	Video manipulation that replaces one person's face with another's. This is the main stimulus type in the DFDC dataset.
Voice clone	Synthetic speech that imitates a specific person's voice.
TTS	Text-to-speech synthesis. It is not necessarily speaker-specific.
GAN	Generative adversarial network, the model family behind StyleGAN.
StyleGAN2 / StyleGAN3	NVIDIA GAN architectures for high-resolution faces. StyleGAN2 is the most common stimulus source in 2021–2023 perception studies.
Diffusion model	A generative model family (for example Stable Diffusion or Midjourney) that has largely replaced GANs since 2022.
DFDC	Deepfake Detection Challenge (Meta/Facebook, 2019–2020), a dataset and competition of face-swap videos.
FFHQ	Flickr-Faces-HQ, a 70,000-image face dataset used to train StyleGAN and as a source of real comparison faces.
AI hyperrealism	The finding that AI-generated faces are judged human more often than real human faces (Miller et al., 2023).
Uncanny valley	Mori's (1970) hypothesis that affinity for humanlike entities drops sharply just short of full human likeness.
Perceptual mismatch	The hypothesis that eeriness comes from inconsistent realism across features, such as realistic eyes on an artificial face.
Atypicality	Deviation from typical human proportions or features, which is penalised more at high realism.
Categorization ambiguity	The hypothesis that uncertainty about whether something is human or non-human causes eeriness. It is weakly supported.
Uncanny valley of mind	Unease caused by attributing feelings or experience to machines, independent of how they look.
Dehumanization hypothesis	The proposal that agents are first processed as human, and that detecting non-human cues then triggers devaluation.
Predictive coding	The theory that perception compares sensory input against predictions. The prediction error from a mismatch is proposed to underlie the uncanny response.
vmPFC	Ventromedial prefrontal cortex, a region involved in valuation. Its activity tracks the uncanny valley's likability profile.
N170 / EPN / LPP	Event-related potential components: N170 (about 170 ms) reflects early face encoding; EPN (early posterior negativity) and LPP (late positive potential) reflect emotional and attentional processing.
d′ (d-prime)	Signal-detection sensitivity. 0 means chance; higher means better discrimination.
β / criterion	Signal-detection response bias. A liberal "real" bias means a tendency to answer "real".
Super-recogniser	A person with exceptional face-recognition ability, typically in the top ~2% on standard tests.
2AFC	Two-alternative forced choice: picking which of two items is fake.
Chance	50% accuracy in a balanced binary real-vs-fake task.
C2PA	Coalition for Content Provenance and Authenticity, a standard for cryptographically signed content credentials.