Deepfake

Working Definition

A deepfake is an image, audio recording, or video generated or manipulated by deep-learning models so as to depict an identifiable, real person saying or doing something they never said or did, in a manner that appears authentic to an ordinary observer. Three elements define the concept: identity impersonation or synthesis, the appearance of documentary authenticity, and automated neural-network generation (autoencoders, generative adversarial networks, diffusion models, or neural voice cloning). In loose everyday usage, the word is often applied to any AI-generated image; in its strict technical and legal definition—codified in Article 3(60) of the European Union’s 2024 AI Act—it designates AI-generated or manipulated image, audio, or video content that “resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.” Low-tech edits that require no machine learning (cropping, slowing down speech to mimic intoxication, or attaching a false caption to genuine footage) are called cheap fakes and belong to a different analytical category, even though they often deceive larger audiences.

Genesis and Global Lineage

The technique is older than the name. Automated facial reenactment synced to an audio track was demonstrated in academic computer graphics as early as the Video Rewrite system (1997), followed by the invention of generative adversarial networks (GANs) by Ian Goodfellow and colleagues in June 2014 and real-time facial transfer projects such as Face2Face (2016) and Synthesizing Obama (2017). The word deepfake itself was born on Reddit: on December 11, 2017, journalist Samantha Cole reported in Motherboard (Vice) that an anonymous Reddit user posting under the handle “deepfakes” (a portmanteau of deep learning and fake) was using open-source libraries (TensorFlow, Keras) and publicly scraped photographs to swap the faces of female celebrities onto pornographic videos. The user’s handle quickly became the generic noun for the medium. Over the next three years, the concept was captured by two influential policy circuits: a legal and national-security literature anchored by Bobby Chesney and Danielle Citron’s 2019 California Law Review article (circulated in draft in 2018), which broadened the term to cover all hyper-realistic digital falsification and coined the concept of the liar’s dividend, and a geopolitical commentary wave exemplified by Nina Schick’s Deep Fakes and the Infocalypse (2020). While elite commentary fixated on scenarios of televised election sabotage, empirical audits—beginning with Deeptrace (later Sensity) in 2019, which found that 96% of all deepfake videos online were non-consensual pornography targeting women—showed that the primary real-world harm was gendered sexual abuse, a pattern underscored again in January 2024 when viral deepfake pornography of Taylor Swift finally prompted legislative action.

Central European Lineage

While the technology originated in North American labs, its most famous wartime test case—and one of its sharpest electoral tests—took place in Central and Eastern Europe. On March 16, 2022, three weeks into Russia’s full-scale invasion of Ukraine, hackers compromised the live news ticker of the Ukrainian television channel Ukraine 24 and circulated a synthetic video across social media and pro-Kremlin Telegram channels in which President Volodymyr Zelenskyy appeared to tell Ukrainian citizens that he had fled Kyiv and that soldiers should lay down their weapons. As the Atlantic Council’s DFRLab documented, the video was technically crude—Zelenskyy’s head proportions, lighting, and accent were visibly off—and Zelenskyy neutralized it within hours by posting a real smartphone video from his Kyiv office dismissing the “childish provocation.” Crucially, Ukrainian military intelligence had spent weeks “pre-bunking” the tactic by warning the public in advance that Russia was preparing deepfake surrender videos, turning the March 16 incident into a textbook case study (analyzed by John Twomey et al. in PLOS ONE, 2023) of how institutional preparedness defeats synthetic media. By contrast, in September 2023 in neighboring Slovakia, a fabricated AI audio clip purporting to show liberal leader Michal Šimečka discussing vote-rigging and hiking beer prices was released on social media during the country’s statutory 48-hour pre-election silence period (cisza wyborcza), when media outlets and campaigns were legally barred from publishing rebuttals—demonstrating how local electoral rules can inadvertently amplify a deepfake’s timing advantage.

Why Now

Three thresholds were crossed between 2018 and 2026. First, the technical barrier collapsed: latent diffusion models (2022), commercial video generators, and zero-shot voice cloning requiring only three seconds of audio (2023) turned a week-long GPU project into a one-click smartphone feature. Second, the global “super-election year” of 2024 tested political deepfakes in the wild—from a robocall impersonating US President Joe Biden in New Hampshire to synthetic campaign clips in India, Indonesia, and Europe—yielding a paradoxical verdict: while voice clones and cheap fakes caused local disruption, the apocalyptic electoral collapse predicted in 2020 did not materialize. Third, legal and financial stakes escalated: the EU’s AI Act (2024) introduced mandatory disclosure and watermarking obligations for deepfakes (Article 50, applicable August 2, 2026), while the US, UK, and South Korea criminalized non-consensual deepfake pornography; meanwhile, a February 2024 Hong Kong heist using a live deepfake video call of a CFO to steal $25 million proved that corporate fraud is a far more lucrative market for deepfakes than election interference.

Explanatory Power

The concept exposes that audiovisual recordings were never self-authenticating “raw facts,” but a fragile social institution whose credibility rested on the high cost of forgery. Three mechanisms stand out. First, Chesney and Citron’s liar’s dividend: once the public knows realistic fake videos exist, wrongdoers caught on genuine tape can plausibly claim “that’s an AI deepfake,” meaning public-awareness campaigns about deepfakes paradoxically strengthen the shield available to liars. Second, the asymmetry of harm: in politics, a deepfake primarily sows momentary doubt, whereas in private life—where over 90% of deepfakes target women and schoolgirls with non-consensual sexual imagery—the harm is immediate humiliation that no forensic “debunking” can undo. Third, the adversarial economics of detection: because every forensic detector can be used as a training loss function to make the next generator undetectable, purely algorithmic detection is a losing race; durable defense depends on institutional provenance (cryptographic capture standards like C2PA, trusted channels, and pre-bunking).

Axes of Debate

Underappreciated Angle

The most enduring political casualty of the deepfake era may not be the gullible citizen who believes a lie, but the whistleblower or war-crimes witness who can no longer prove a truth. For a century, democratic accountability—from Watergate tapes and human-rights documentation to bystander videos of state violence—relied on a cheap, decentralized presumption of authenticity: an ordinary citizen with a camera could challenge a government minister. As synthetic generation drives the cost of plausible forgery toward zero, that evidentiary subsidy flips. Proving a recording is genuine now requires forensic chain-of-custody logs, cryptographic signatures, and institutional corroboration—resources monopolized by intelligence agencies, wealthy corporations, and legacy newsrooms. The “liar’s dividend” is thus also an incumbent’s dividend: generalized skepticism toward raw audiovisual evidence redistributes epistemic power away from lone witnesses and back toward gatekeeping institutions that can afford to certify reality.