WITNESS Raises Concerns to Meta Oversight Board on Generative AI in Elections
On 7 May 2026, WITNESS submitted a public comment to Meta’s Oversight Board, an independent body that reviews Meta’s content moderation decisions, in response to a case involving an AI-generated video of a prominent Hungarian politician posted ahead of Hungary’s recent elections. Although the Oversight Board is no longer reviewing this case as the video has since been removed, WITNESS submission provides valuable insights on the use of generative AI in the electoral contexts.
The case centered on an eight-second video appearing to depict Péter Magyar, now the Prime Minister of Hungary and then an opposition candidate, expressing frustration over the use of robocalls in political campaigns. Posted on Facebook in November 2025, when Magyar was still the country’s main opposition leader, the video appeared to be AI-generated and raised broader concerns about how synthetic media can shape electoral discourse and distort the information environment.
The content circulated during the 2026 Hungarian elections highlighted the escalating challenge posed by synthetic media in high-stakes political contexts, where AI was used to amplify inflammatory narratives within an increasingly saturated information environment. By weaponizing geopolitical anxieties surrounding Russia’s invasion of Ukraine, political campaigns deployed AI-generated content to appeal to voters’ fears, including claims that one party’s victory would lead Hungarians into war.
This highly emotional content coexisted with satire, parody, and low-quality “AI slop,” featuring caricatures of candidates alongside Western celebrities, cartoon characters, and even animated vegetables endorsing political messages. The result was a blended information environment in which authentic, synthetic, humorous, and misleading content reinforced one another, making it increasingly difficult for audiences to assess credibility and distinguish fact from fiction.
Importantly, the risks posed by generative AI in elections are no longer limited to whether individual pieces of content are believable or technically deceptive. AI-generated media increasingly shapes political discourse through scale, repetition, emotional manipulation, and coordinated amplification across platforms and networks of accounts. Even clearly labeled or obviously synthetic content can still influence perceptions, reinforce narratives, and contribute to confusion, distrust, and polarization.
“Overall, the role of AI in elections is best understood not as a problem of isolated deceptive artifacts, but as a structural transformation of the information ecosystem. AI-generated content amplifies polarization, enables scalable and often opaque influence operations (including potential foreign interference), facilitates harassment and targeted attacks, and erodes the shared baseline of trust that democratic processes depend on,” said shirin anlen, AI Research Technologist and Impact Manager at WITNESS.
In its submission, WITNESS highlights three key dynamics:
- The scale, quality, and volume of content: AI significantly lowers the cost and barrier to producing persuasive, multimodal political content at scale. This enables the rapid creation and dissemination of large volumes of content capable of amplifying political narratives, suppressing voter participation, and harassing public figures across platforms.
Women politicians are disproportionately targeted, particularly through non-consensual intimate imagery (NCII). More broadly, the most impactful content is often not the most technically sophisticated, but the most emotionally resonant — leveraging fear, identity, outrage, and political division to shape public opinion and engagement.
- The erosion of trust: The growing volume of synthetic content makes it increasingly difficult for users to distinguish between authentic and manipulated media. At the same time, this environment enables “plausible deniability,” where political actors dismiss authentic footage or recordings as AI-generated.
WITNESS’ Deepfake Rapid Response Force (DRRF) has found that approximately one-third of analyzed cases involved authentic media being falsely labeled as AI-generated. This dynamic undermines trust not only in individual pieces of content, but also in audiovisual evidence more broadly, weakening journalism, accountability, and democratic processes that rely on shared facts.
- The scalability and subtlety of influence operations: AI enhances the scalability and adaptability of influence operations by enabling political actors and aligned networks to rapidly generate content tailored to specific narratives, emotions, and audiences. Rather than relying solely on highly realistic deepfakes, these campaigns often deploy large volumes of synthetic or partially synthetic content that gradually shapes perceptions over time.
The Hungarian elections demonstrated how AI-generated content can reinforce geopolitical narratives and normalize political messaging through repetition and emotional storytelling. Increasingly, politically aligned content creators and AI-generated influencer personas operate as interconnected influence ecosystems, amplifying narratives through cross-posting, algorithmic engagement, and manufactured authenticity.
These dynamics create a persistent form of narrative shaping that can distort perceptions of public opinion, manufacture a false sense of consensus, and influence voter attitudes without relying on any single piece of clearly deceptive or policy-violating content.
WITNESS’ submission argues that platform responses focused primarily on labeling or detecting individual pieces of AI-generated content are insufficient to address these broader systemic risks. Addressing these challenges requires stronger provenance and authenticity infrastructure, including wider adoption of standards such as C2PA; more rigorous socio-technical evaluation of detection and moderation systems, including through initiatives such as the TRIED Benchmark; and greater attention to coordinated networks, cross-platform amplification, and the limitations of current detection tools, particularly in under-resourced languages and contexts.