# Deep Fakes A portmanteau of "deep learning" and "fake," coined in late 2017 by an eponymous Reddit user who used face-swap models to graft celebrity faces onto pornographic video. The technique was originally a non-consensual-imagery weapon; the name stuck and rapidly generalized to cover any AI-synthesized media of a specific real person. By the early 2020s, generative-AI advances (StyleGAN, diffusion models, voice-cloning systems) had pushed deep-fake capabilities from a hobbyist subreddit to a baseline disinformation tool available at consumer pricing. **Underlying technology.** Deepfakes are an application of [[Generative AI]]. The 2017 face-swap wave was driven by the Generative Adversarial Network (a generator and discriminator trained against each other); today's higher-fidelity image and video synthesis increasingly relies on the Diffusion Model architecture, alongside dedicated voice-cloning systems. The same generative stack that powers creative tooling powers the impersonation — which is why "deepfake" and "synthetic media" name the malicious and neutral faces of one technology. ## The naming history matters Two things are notable about the etymology that bear on how the term operates: 1. **Origin in non-consensual sexual imagery, not politics.** The first wave of public deep-fakes was face-swap pornography of female celebrities. This is the term's original referent and remains a substantial fraction of all deep-fake content produced. Political deep-fakes are downstream of the technique developed for harassment. 2. **Pseudonymous coinage by the practitioner.** Unlike most disinformation terminology — coined by academics or journalists *describing* a phenomenon — "deep fakes" was coined by the people *making* them, as a casual self-label that journalists then adopted. The provenance is one reason the term carries a slight tabloid coloration that "synthetic media" (the more academic alternative) does not. ## What the term covers (and overstates) "Deep fakes" has come to function as a catch-all for synthetic media of named individuals, but the analytical category is murkier than the catch-all suggests: - **Face-swaps** — replacing one person's face with another in video. The original technique. - **Lip-sync / mouth-reanimation** — driving a target person's mouth from a different audio track. The most-deployable misinformation form. - **Voice clones** — short audio samples regenerated as full speech in the target's voice. Increasingly indistinguishable from authentic recordings as of 2024–2026. - **Text-to-video of named figures** — Sora and successors. Currently weakest in fine motor accuracy but improving fast. - **Text-to-image of named figures** — Midjourney, FLUX, Imagen. Already at near-photorealistic quality. When journalists call something a "deep fake," they often mean any of the above. The umbrella usage obscures meaningful differences in production cost, detection difficulty, and democratic impact. ## The political-disinformation angle Several deep-fake-driven political incidents have entered the record: - **The Biden-robocall (January 2024, New Hampshire primary)** — voice-cloned [[Joe Biden]] told New Hampshire Democrats not to vote in the primary. Traced to Steve Kramer (a campaign consultant working with the Dean Phillips campaign). $6M FCC fine; Kramer faced criminal charges. The first deep-fake election-interference operation prosecuted in the US. - **The Pope-in-Balenciaga image (March 2023)** — Midjourney synthesis of Pope Francis in a white Balenciaga puffer jacket. Not malicious in intent but the "viral image proves the public can't tell" benchmark. - **The Pentagon-explosion image (May 2023)** — viral fake image of an explosion near the Pentagon caused brief stock-market dip; demonstrated AI-image impact on financial systems. - **[[Volodymyr Zelensky]] surrender video (March 2022)** — early deep-fake attempting to undermine Ukrainian morale. Crude execution, quickly debunked, but a doctrinal preview. ## "The liar's dividend" Law professors Bobby Chesney and Danielle Citron coined this companion term in their 2019 paper *Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security*. The dividend: as deep fakes become plausible, real audio and video become deniable. A politician caught on tape can claim it's synthetic. The corrosion of evidence runs in both directions — fakery becomes plausible *and* real evidence becomes contestable. The dividend may matter more politically than any specific deep-fake operation. ## Parody, satire, and the law The legal line between a protected parody and an actionable fake comes from [[Hustler Magazine v. Falwell]] (1988): content that no reasonable viewer would take as a factual claim is protected. Realistic synthetic video tests that line. California's AB 2839, passed after [[Elon Musk]] reposted an AI-voiced parody of Kamala Harris in July 2024, was struck down on August 29, 2025; the judge wrote that "a mandatory disclaimer for parody or satire would kill the joke." ## See Also - [[Disinformation]] — the broader phenomenon - [[Fake News]] — adjacent rhetorical category (linguistic rather than visual) - [[Disinformation Hub]] — Subject hub - [[Information Warfare]] — operational-domain hub (deep-fakes are a deployable weapon there) - [[Pentagon Influence Operations - gc_ Generation]] — At least one gc\_ site used an AI-generated newscaster on [[Instagram]] to present content, demonstrating state-level adoption of synthetic-media personas for influence operations - [[Generative AI]] · Diffusion Model · Generative Adversarial Network — the underlying technology --- *Merged from a former `/Concepts/Deepfake.md` stub (2026-06-22): this page is the canonical entry for **Deepfake / Deep Fakes / synthetic media**. The "Deepfake" one-word title now resolves here via alias.*