Synthetic media

Working Definition

Synthetic media refers to images, audio, video, and multimodal artifacts generated or substantially modified by machine-learning algorithms rather than captured by a physical lens or microphone recording an event that actually occurred in front of it. The category is intentionally broader than deepfake: while a deepfake implies deceptive or non-consensual impersonation of a real person, synthetic media encompasses everything from malicious voice clones and political hoaxes to openly labeled virtual news anchors, commercial dubbing, and AI cinema. In technical usage, the term simply describes a generative production pipeline; in media theory and political analysis—following Robert Chesney, Danielle Citron, and Nina Schick—it names a historical shift in which audiovisual recordings lose their default nineteenth- and twentieth-century status as indexical proof of reality, shifting the burden of verification from the sensory artifact itself onto its cryptographic and institutional provenance.

Genesis and Global Lineage

The term began circulating among computer-graphics and ML researchers around 2017, as Generative Adversarial Networks (Ian Goodfellow, 2014) produced photorealistic faces and the first face-swapping deepfakes appeared on Reddit. Its legal and political framework was established by Robert Chesney and Danielle Keats Citron in Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security (2018 preprint; California Law Review, 2019), which cataloged individual and democratic harms and coined the liar’s dividend: as the public learns how easy it is to fabricate video and audio, guilty politicians can dismiss authentic recordings of corruption as “just an AI fake.” In 2019, Britt Paris and Joan Donovan (Deepfakes and Cheap Fakes, Data & Society) showed that low-tech edits—cropping, slowing video to make a speaker look intoxicated, or false captions—still caused most real-world manipulation, reminding analysts that recordings have always required institutional “social work” to count as evidence. In 2020, Nina Schick’s Deepfakes: The Coming Infocalypse popularized synthetic media as the umbrella term, which AI avatar and dubbing startups after 2022 adopted as a sanitized industry label.

Central European Lineage

Central Europe connects to the concept through both a philosophical anticipation from postwar Kraków and the flagship EU test case of electoral AI manipulation. In Chapter VI of Summa technologiae (Wydawnictwo Literackie, Kraków 1964), the Lwów- and Kraków-educated writer and philosopher Stanisław Lem coined phantomatics (fantomatyka): generating synthetic sensory inputs indistinguishable from physical reality. Asking whether “perfect illusions” are still illusions, Lem diagnosed a “solipsistic situation” in which interpersonal trust collapses because no citizen can tell whether a message is authentic or simulated—anticipating the epistemic core of the synthetic-media debate. Sixty years later, Bratislava supplied the empirical benchmark known internationally as “the Slovak case.” On September 28, 2023—during Slovakia’s mandatory 48-hour pre-election media moratorium ahead of the September 30 parliamentary vote—a fabricated audio clip went viral on Telegram and Facebook purporting to show Michal Šimečka, leader of Progressive Slovakia (Progresívne Slovensko), and journalist Monika Tódová (Denník N) conspiring to buy votes from the Roma minority. Documented by the International Press Institute as one of the first major EU cases of synthetic audio weaponized against a journalist during an election moratorium, it was analyzed by Lluis de Nadal and Peter Jančárik (HKS Misinformation Review, 2024) as the defining test case for separating real electoral vulnerability from “deepfake hype.”

Why Now

Between 2022 and 2026, three thresholds were crossed simultaneously. First, diffusion image/video generators (DALL·E 2, Midjourney, Stable Diffusion in 2022; Sora and rivals in 2024–2025) and zero-shot voice-cloning tools mimicking any voice from five seconds of audio drove the cost of convincing audiovisual synthesis near zero. Second, incidents moved into elections, courts, and schools: the September 2023 Slovak audio hoax, January 2024 robocalls impersonating President Joe Biden in New Hampshire, multi-million-dollar corporate wire frauds via live deepfake video calls, and an epidemic of non-consensual synthetic intimate imagery targeting women and girls worldwide. Third, regulation and cryptographic standards arrived: Article 50 of the EU AI Act (2024) mandated machine-readable marking of synthetic content and disclosure of deepfakes, while industry coalitions led by Adobe, Microsoft, and the BBC rolled out the C2PA Content Credentials standard. Meanwhile, passive forensic detection continued to lag: during the 2023 Slovak hoax, forensics labs struggled to prove with certainty which model generated the compressed social-media audio within the 48-hour electoral window.

Explanatory Power

The deepest impact of generative media is not that everyone believes every fake video, but that the default epistemic status of all audiovisual recording is inverted. First, it explains Chesney and Citron’s liar’s dividend: since 2023, politicians and defendants caught on authentic dashcam or wiretap recordings routinely claim the footage is AI-generated, forcing courts and journalists to reconstruct an unbroken chain of custody before a tape is believed. Second, it explains why watermarking mandates (EU AI Act Article 50) operate asymmetrically: law-abiding studios attach C2PA metadata and “AI-generated” badges, whereas intelligence services, scammers, and harassment networks strip metadata or run local open-weight models without watermarks—meaning labels prevent accidental confusion among honest actors, but never stop deliberate forgery. Third, it reveals a shift in media power from the artifact back to the institution: in the twentieth century, amateur footage from an unknown bystander could challenge state narratives by its visual indexicality alone; in an environment saturated with synthetic media, an unsigned video carries less weight, shifting authority back toward newsrooms, state agencies, and hardware vendors controlling cryptographic keys.

Axes of Debate

Underappreciated Angle

The most corrosive long-term effect of synthetic media is not that people will believe everything they see, but that they gain a license to disbelieve whatever they find uncomfortable. When Walter Benjamin wrote in 1936 about mechanical reproduction, the photograph still retained what Charles Sanders Peirce called indexicality: like a footprint in mud, a chemical photo was a physical trace of light bouncing off a real object. Statistical synthesis severs indexicality altogether: a diffusion image is a draw from a probability distribution that never needed a physical referent. Yet as Paris and Donovan observe, raw indexicality was never enough on its own; a photo counted as evidence because social institutions—photojournalists, editors, archivists, and courts enforcing perjury laws—performed the social work of vouching for it. This returns us to the question Lem posed in 1964: once sensory experience can be synthesized on demand, truth becomes a question of which human institutions have earned enough costly trust to vouch for reality.