The Deepfake Was Never the Real Problem. It Was Always What Deepfakes Do to Public Trust.

Eight months ago, Prof. Cary Coglianese of the Penn Program on Regulation warned that the most consequential effect of AI-generated election content is not the content itself. Eight months, several state elections, and one general election cycle later, the warning has become sharper.

Eight months ago, A4G recorded a conversation with Prof. Cary Coglianese, Edward B. Shils Professor of Law at the University of Pennsylvania and Director of the Penn Program on Regulation, for the third episode of the Governance Dialogues series.

One thing he said about AI and elections has not, since, been repeated as often in the Indian public conversation as it deserves to be.

“The most dangerous deepfake isn’t the video — it’s the voter who stops trusting any information at all.” — Prof. Cary Coglianese

The point is not the one the public conversation about deepfakes has typically landed on.

The dominant framing has been that deepfakes are dangerous because they trick people. A fabricated video of a political leader circulates on WhatsApp or X. Some voters see it. Some believe it. Some update their voting intention as a result. Aggregate the individual mistakes and the result is a distortion of the election.

That concern is real and well-documented. It is not the largest concern.

The larger concern, as Coglianese framed it, is what happens after enough deepfakes have circulated in a given electoral environment for voters to stop treating video, audio, and image content as reliable evidence about the world. Once the epistemic ground has been disturbed enough that a real recording can be plausibly dismissed as fabricated, the deployment of AI-generated content has done its most consequential damage — not through the fakes it succeeded in placing, but through the doubt it succeeded in producing about all content.

The liar’s dividend

The phenomenon has a name in legal scholarship. The liar’s dividend. It was introduced in a widely cited 2019 paper by Bobby Chesney and Danielle Citron published in the California Law Review — the observation that public actors gain a defensive advantage in an environment saturated with deepfakes, because they can now credibly dismiss any inconvenient real evidence as fabricated. The liar wins twice. Once through the fakes that succeed in circulating. And again through the doubt those fakes cast on everything else.

The Chesney-Citron paper was written at a time when convincing deepfakes required professional technical resources and significant time. Six years later, both requirements have collapsed. Consumer-grade tools produce convincing synthetic video and audio in minutes. The Indian language coverage of these tools has expanded substantially. Voice cloning, image manipulation, and video synthesis in Hindi, Tamil, Telugu, Bengali, Marathi, and multiple other Indian languages is now trivially accessible.

Applied to the Indian context

The theoretical concern moves to operational.

The 2024 Indian general election was the first Indian national election in which AI-generated content circulated at scale in multiple Indian languages, targeting specific voter groups through WhatsApp and other messaging platforms. Fact-checking organisations including Alt News, Boom, and The Quint documented cases in which deepfake videos moved through language- and region-specific networks faster than any single fact-checking operation could evaluate them.

The 2026 and 2027 state elections are now being fought on infrastructure that is significantly more sophisticated. The 2029 general election will be fought on infrastructure that does not yet exist in publicly visible form, but is being built right now.

The direct effect of these deployments — voters being persuaded by fabricated content — has begun to be tracked, imperfectly, by academic researchers and civil-society organisations. The Election Commission of India has issued guidance. The IT Rules have been amended to address synthetic content in electoral contexts. These are real responses.

The indirect effect — the erosion of shared evidential trust — is significantly harder to measure and, by every indication from research on comparable environments internationally, substantially more consequential. What does it mean, in operational terms, to run a democracy when a growing share of the electorate no longer treats a recording as evidence of what it depicts? What replaces the recording as the basis for shared public knowledge about political actors and their statements? Whose word gets trusted then, and on what basis?

These are not questions the Indian electoral system will face at some point in the distant future. They are questions the system is being handed right now, without a coherent institutional framework for answering them.

What Coglianese was actually warning about

His point, made eight months ago and sharpened by every intervening election cycle in every major democracy, is that the deepfake was never the real problem. It was always what deepfakes do to the possibility of shared public evidence at all. That is a structural harm to democratic decision-making that the current Indian regulatory response is not yet adequately addressing — because the response has been built around the direct question of removing individual pieces of fabricated content, and the harder question is what a country does when public trust in evidence itself begins to erode.

A serious response would engage the second-order problem directly. Not just tools for identifying fakes. Not just penalties for creators. Institutional investment in trusted intermediaries — public broadcasters, independent electoral bodies, verified journalism organisations — whose reliability under stress is worth defending because those institutions are what maintain the shared evidential floor a democracy requires.

The current Indian conversation has not, on the whole, begun to engage this level of the problem. It should.

The A4G Intelex Essay Contest 2026 closes on July 10. Two days from now. The theme is AI, Work and Society. If the collapse of shared electoral evidence in the AI era is the question your work, your research, or your ground-level exposure has been circling, this is the moment to put your argument on the record. The jury includes Anil K. Antony, Dr. Prateek Sharma (Vice Chancellor, Delhi Technological University), Shahed Arora, and Dr. Nandini Chatterjee Singh (UNESCO MGIEP).

Learn more about A4G → **a4gcollab.org**