Elections and synthetic media

Lesson 1 of 5 in AI, Elections, Labor, Competition, and the Planet.

Two days before Slovakia’s September 2023 election — inside the legal quiet period when candidates could no longer respond — an audio clip spread of liberal party leader Michal Šimečka apparently discussing how to rig the vote. It was an AI fabrication. He lost. Whether the clip changed the result is unprovable, and that unprovability is itself the lesson: synthetic media attacks arrive faster than rebuttal, timed to exploit the exact rules (moratoria, ad blackouts) designed to protect elections.

Four months later, New Hampshire Democratic-primary voters received robocalls in President Biden’s cloned voice telling them not to vote. The response chain is a miniature of the whole US approach: the FCC ruled within weeks that AI-generated voices in robocalls fall under the existing Telephone Consumer Protection Act, the political consultant behind the calls was criminally charged and fined millions, and the carrier that transmitted them settled for $1 million. Note what did that work: a 1991 statute, stretched. The US governs election deepfakes mostly by retrofitting old law — because new federal law never passed.

The state legislatures moved instead. Texas criminalised election deepfakes in 2019, California followed, and after 2023 a wave took the count to roughly half the states. Nearly all share one architecture: they target synthetic depictions of candidates within a window before an election (30–90 days), and they split between requiring a disclosure label and prohibiting the content outright, with carve-outs for satire and news. Then came the constitutional collision: California’s aggressive AB 2839 — allowing candidates to sue over materially deceptive election content — was enjoined within weeks in Kohls v. Bonta on First Amendment grounds. The stable US equilibrium, for now: disclosure mandates survive scrutiny more easily than bans, and parody exceptions are load-bearing.

The EU built the opposite architecture — platform-side and systemic. Under the Digital Services Act, very large online platforms must assess and mitigate systemic risks to electoral processes (Articles 34–35), with 2024 Commission election guidelines telling them how, including labeling AI-generated content. The AI Act’s Article 50(4) requires deepfake disclosure by deployers, and Annex III makes AI systems intended to influence elections high-risk. A third instrument completes the set: the Political Advertising Regulation (2024/900, applying from October 2025) requires every political ad to carry a transparency label and bans targeting based on sensitive personal data. Where the US regulates the lie about the candidate, the EU regulates the infrastructure that carries it.

Key terms: synthetic media, deepfake, dsa, liar’s dividend, political advertising

Interactive sorting exercise: Sort each instrument by its regulatory strategy: does it mandate a label, prohibit content, or impose duties on the platform/infrastructure layer?

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.