Seeing Is No Longer Believing: Regulating Deepfakes

Introduction: The Deepfake Threat to Democracy

In October 2023, a deepfake audio clip of Slovak opposition leader Michal Šimečka circulated on social media two days before the country's national election. The clip, fabricated using AI voice synthesis, depicted Šimečka apparently discussing how to rig the vote. Though debunked by fact-checkers, it spread widely during the mandatory pre-election media silence, a period when candidates cannot legally respond to attacks. Šimečka's party lost. Whether the deepfake decided the outcome is unknowable; that it poisoned the information environment at the worst possible moment is not. That episode is not an anomaly. It is a preview. As AI-generated synthetic media grows more sophisticated and accessible, democratic societies face a dilemma that cannot be resolved by market forces or voluntary platform policies alone. Deepfakes and AI-generated misinformation pose a sufficient threat to democratic discourse to warrant regulatory intervention because they systematically undermine the epistemic conditions on which free, informed political participation depends — and the window to act before these harms become irreversible is narrowing.

Why Deepfakes Are Categorically Different

To understand why deepfakes are categorically different from older forms of political deception, it helps to understand what makes democratic discourse function in the first place. Political theorists broadly agree that legitimate democratic outcomes require what Jürgen Habermas described as the conditions of communicative rationality: participants must be able to evaluate claims, weigh evidence, and form genuine preferences rather than manufactured ones (Habermas 25). Propaganda and spin have always distorted this ideal, but they operated within a world where citizens retained a baseline epistemic anchor — the assumption that photographs and audio recordings constituted evidence. Deepfakes dissolve that anchor. They do not merely add false claims to the information environment; they corrode the mechanism by which citizens distinguish true claims from false ones. When any video of a candidate confessing to a crime, any audio of a general ordering an illegal strike, or any image of a diplomat meeting with an adversary can be plausibly fabricated, the rational response is to doubt everything. Scholars call this the "liar's dividend": the ability to dismiss authentic, damaging evidence by claiming it might be synthetic (Chesney and Citron 1775). The damage deepfakes cause is not only the false beliefs they create but the generalized skepticism they license.

The Scale and Trajectory of the Problem

The scale and trajectory of this problem give the liar's dividend real teeth. The volume of synthetic media is growing at a pace that outstrips detection capacity. A 2023 report by the RAND Corporation on AI and information operations found that generative AI tools have dramatically lowered the cost and technical skill required to produce convincing synthetic media, making large-scale influence campaigns accessible to actors who previously lacked the resources to mount them. Research tracking deepfake video online found that the number of detected deepfake videos doubled roughly every six months between 2019 and 2023 (Vaccari and Chadwick 4). Detection technology, meanwhile, consistently lags behind generation technology — an asymmetry that is structurally baked into the problem, because detecting a synthetic artifact requires knowing what artifacts a new generation technique leaves behind, information that only becomes available after the technique is deployed. Platform content moderation has not closed this gap. A Pew Research Center survey conducted in 2023 found that 79 percent of Americans were concerned about the use of AI to create misleading content, while confidence in platforms' ability to police such content remained low. The problem is real, growing, and not being solved by existing voluntary mechanisms.

The Free Speech Counterargument

Skeptics of regulation often invoke free speech as a near-absolute barrier to government intervention in political expression, and this objection deserves serious engagement before it can be answered. The strongest version of the free speech argument runs as follows: the history of government attempts to regulate political speech is a history of abuse, because governments have an obvious self-interest in silencing speech that challenges their authority. Any regulatory framework broad enough to catch harmful deepfakes will be broad enough to be weaponized against satire, parody, legitimate whistleblowing, or dissenting political expression that powerful actors find inconvenient. Proponents of this view point to the First Amendment doctrine that political speech receives the highest constitutional protection precisely because the democratic process is itself the intended remedy for political disagreement. On this reading, the cure — government control over political expression — is worse than the disease. Advocates like legal scholar Eugene Volokh have argued that AI-generated content is simply speech and that content-based restrictions face an inherently high burden of justification under existing First Amendment jurisprudence.

What Effective Regulation Actually Looks Like

This argument is thoughtful, but it proves too much and rests on an assumption that the regulatory proposals actually on the table do not support. First, the free speech objection conflates restricting what can be said with regulating whether content must be disclosed as synthetic. The strongest and most legally durable regulatory proposals do not ban deepfakes; they mandate disclosure. A law requiring that AI-generated political advertising carry a clear label — "this content was produced using artificial intelligence" — does not prevent anyone from making or distributing synthetic media. It prevents them from passing it off as authentic documentary evidence. This is no different in principle from existing truth-in-advertising rules or securities disclosure requirements, neither of which has been seriously treated as an unconstitutional abridgment of free expression. Second, the free speech argument implicitly assumes that the current unregulated environment is neutral with respect to speech. It is not. Deepfakes selectively silence the speech of the people they target: a fabricated video depicting a candidate committing a crime does not add to democratic debate, it forecloses the ability of real voters to evaluate real candidates. Protecting the conditions of democratic discourse is not the enemy of free speech — it is, as Chesney and Citron argue, the precondition for speech to function democratically at all (Chesney and Citron 1801).

Concerns about regulatory overreach and innovation chilling are also addressable through the specific architecture of proposed policies. Effective regulatory intervention in this space does not require a single sweeping prohibition; it requires a layered approach targeting the most harmful uses while preserving space for creative and technological development. The first and most urgent measure is mandatory disclosure for AI-generated political content. The European Union's AI Act, which took effect in 2024, includes provisions requiring that synthetic media be labeled at the point of generation, and several U.S. states, including California and Texas, have enacted statutes specifically targeting deepfakes in political advertising. These laws provide a template: narrowly scoped to electoral contexts, they avoid the overbreadth problem that would threaten legitimate creative uses. The second layer of intervention involves platform liability reform. Under the current Section 230 framework in the United States, platforms bear no legal responsibility for failing to remove demonstrably synthetic content even when they have been notified of its existence and electoral impact. Amending Section 230 to create a limited exception for verified AI-generated electoral disinformation — one that requires platforms to act on documented reports rather than imposing proactive screening obligations — would align incentives without transforming platforms into state censors. Third, investment in public-facing detection infrastructure, funded by levies on AI developers, would democratize access to tools that currently exist only in well-funded research labs, giving journalists and civil society organizations the capacity to identify synthetic media in real time (Benkler et al. 12).

The innovation objection — that regulation will stifle the enormous beneficial potential of generative AI — should not be dismissed, but it should be kept in proportion. Disclosure requirements impose trivial compliance costs on legitimate AI developers; the primary burden falls on bad actors who want to disguise synthetic content as real. The broader AI industry's growth has not been hampered by the EU's AI Act, which was negotiated with extensive industry participation and built in risk-tiered obligations that reserve the heaviest regulatory burdens for the highest-risk applications. The same logic applies here: labeling a political advertisement as AI-generated is not a significant barrier to innovation. It is a basic condition of honest political communication. Democracies routinely require disclosures that impose costs on speakers — campaign finance reports, pharmaceutical advertising standards, food labeling laws — because the public interest in informed decision-making outweighs the speaker's preference for unmediated messaging. There is no principled reason why synthetic political media should be exempt from that logic (Sunstein 44).

Conclusion: The Cost of Inaction

The stakes of inaction are not abstract. Democratic systems depend on citizens' capacity to form genuine political judgments, and that capacity depends on a minimally trustworthy information environment. As synthetic media becomes cheaper, faster, and more convincing, the window in which disclosure requirements and platform accountability rules can establish norms before the technology normalizes will close. Electoral institutions are slow-moving; they cannot adapt in real time to information environments that are shifting in months, not years. The Slovak election episode illustrates the timing problem with painful clarity: by the time a deepfake is debunked, the election may already be over. Regulatory frameworks that require disclosure at the point of creation, rather than correction after the fact, are the only interventions capable of operating at the speed the problem demands. Waiting for markets or platforms to solve this voluntarily is not a neutral choice; it is a choice to allow the degradation of democratic discourse to continue on its current trajectory. The question is not whether to regulate deepfakes in political contexts, but whether to do so while the institutions being protected are still functional enough to enforce meaningful rules.