Definition
A deepfake is synthetic audiovisual content generated by deep neural networks (originally GAN-based, now predominantly diffusion-based and transformer-conditioned) that depicts an identifiable person saying or doing something they did not say or do. The term, coined on Reddit in 2017, has expanded from face-swap video to encompass voice cloning (≤3 seconds of reference audio suffices for many systems), full-body synthesis, real-time avatar puppeteering, and document forgery (e.g., synthetic passports, medical reports).
Technically, contemporary deepfakes exploit several factorizations: identity disentanglement (separating "who" from "what they're doing"), temporal coherence priors (suppressing flicker artifacts), and alignment to biometric anchors (lip-sync to phonemes, gaze to off-screen referents). Detection has become an arms race: classical artifacts (asymmetric blink rates, inconsistent corneal reflections, frequency-domain noise patterns) are progressively eliminated by each model generation.
Deepfakes constitute an epistemic threat surface distinct from textual disinformation because audiovisual content carries an evolved prima facie credibility — humans encode "seeing is believing" as a deep heuristic. Three secondary effects compound the primary forgery harm: (1) the liar's dividend — genuine recordings can now be plausibly denied as fakes; (2) the chilling effect on testimonial trust generally; (3) non-consensual intimate imagery weaponization, which constitutes the dominant harmful use-case in current empirical surveys (96%+ of in-the-wild deepfakes per Sensity 2019, 2023 follow-ups).
Provenance-based defenses (C2PA content credentials, hardware-anchored capture attestation, watermarking) are being deployed but face fundamental adversarial dynamics — any defense visible to the attacker can be optimized against.
Context
Within the Prismatic Platform, deepfakes are modeled as a first-class threat vector in several subsystems:
- The Decision Engine Source Assessment stage applies provenance-confidence scoring to ingested audiovisual evidence; unsigned media from low-trust channels are downweighted regardless of apparent content.
- NABLA Infinity implements multi-modal cross-checks: claims attested by single-channel evidence (video only, audio only) carry lower epistemic weight than multi-channel corroboration with independent provenance.
- Color Teams Red Team scenarios include deepfake-enabled attack patterns: synthetic-CEO fraud, fabricated political endorsements, false-flag testimonial videos.
- The Kompendium catalogues deepfake-enabled techniques under EA-1.2 (epistemic falsification) with cross-references at /kompendium/07-teoreticke-reflexe/ regarding the philosophical implications for testimonial epistemology.
- Hyperreality — Baudrillard's concept — is increasingly literal under deepfake saturation: simulacra precede and replace the real.
Deepfakes are not merely a "fake content" problem; they are a systemic erosion of audiovisual evidence as a category, with downstream consequences for journalism, law, intelligence, and interpersonal trust.
See Also
- Epistemic Security — Field studying systemic threats to collective cognition
- Info-Ops — Military-grade information operations
- Firehose of Falsehood — Adjacent disinformation technique
- Hyperreality — Baudrillardian frame for synthetic reality
- Delegitimizace — Liar's-dividend dynamics
- Epistemická anihilace — Central kompendium concept
- NABLA Infinity — Platform multi-modal cross-checking
- Kompendium — Full Czech-language taxonomy
Connect & Contribute
Created by Tomáš Korcak (korczis) | Open Source under GHL