The liar's dividend is the phenomenon in which the existence of convincing AI-generated fakes lets bad actors plausibly deny authentic content. The term was introduced by Robert Chesney and Danielle Citron in "Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security," published at 107 California Law Review 1753 (2019) and circulating since 2018 as an SSRN working paper; Chesney is the first-named author. The passage appears in section II.B.2.i, "The Liar's Dividend: Beware the Cry of Deep-Fake News": "Hence what we call the liar's dividend: this dividend flows, perversely, in proportion to success in educating the public about the dangers of deep fakes" (Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security (Chesney & Citron, California Law Review, 2019)). In debates over synthetic-media policy, it is offered as an account of how a small share of fabricated content can degrade trust across an information ecosystem.
Mechanism
As described in the originating article
Chesney and Citron present the liar's dividend as a harm "different in kind" from the others they survey, because it works through denial rather than fabrication: "some of the most dangerous lies take the form of denials." They identify two ways deep fakes help liars deny the truth. In the first, a person accused of wrongdoing fabricates altered audio or video that appears to contradict the accusation, a strategy the authors describe as high-risk except where resource inequality makes exposure unlikely. In the second, the one they name, "liars aiming to dodge responsibility for their real words and actions will become more credible as the public becomes more educated about the threats posed by deep fakes," because "a skeptical public will be primed to doubt the authenticity of real audio and video evidence" (Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security (Chesney & Citron, California Law Review, 2019)).
The authors place the dividend within broader trends of truth skepticism. They cite mounting distrust of traditional news, "stoked relentlessly by President Trump and like-minded sources," and the "fake news" mantra as "a useful substitute for argument when confronted with damaging factual assertions," and anticipate that "fake news" will extend to "deep-fake news." Citing the Edelman Trust Barometer's nine-point fall in the US Trust Index from 2017 to 2018, they argue that the combination of "truth decay and trust decay" creates space for authoritarianism, and that cognitive bias gives motivated reasoners "a convenient excuse" to reject information that is true. In their discussion of platform policy they return to the point, warning that growing awareness of deep fakes invites "bad faith exploitation" through "a well-timed allegation of fakery," and cite it as a reason platforms should give users notice and a chance to challenge removals (Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security (Chesney & Citron, California Law Review, 2019)).
The article concerns manipulated audio and video produced with neural networks and generative adversarial networks, and predates the diffusion image models and large language models to which the term is now applied (Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security (Chesney & Citron, California Law Review, 2019)).
Staged account
A staged version of the dynamic, not found in the originating article, runs as follows. A small proportion of content in an information ecosystem is AI-generated fake. Media coverage amplifies public awareness of that proportion. Users then become systematically skeptical of all content, including authentic content. Public figures exploit this skepticism by claiming that real evidence — video, audio, or documents — is AI-generated. The net effect is that audiences cannot tell what to trust, and the information ecosystem degrades even though direct persuasion by fakes remains rare. This staged account, which runs from the prevalence of fakes to generalized skepticism, is closer to Persily's framing below than to the originating article, which ties the dividend to the success of public education about deep fakes rather than to the share of content that is fake.
Comparison to direct persuasion
Persily's Digitalist Papers essay argues, on the empirical record of the 2023–2024 elections in India, Indonesia, the EU, the UK, and France, that direct persuasion by deepfakes is mostly a "tail problem": the persuadable population exposed to AI-fake political content is small, and most of those exposed are true believers whose votes are not changed. Persily cites Allcott/Gentzkow-style research estimating the "fake news" share of the 2020 US media diet at roughly 0.15%.
By Persily's account, the liar's dividend compounds even at around 1% of content, because a narrow slice of known-fake content erodes confidence in the remaining 99%-plus. He argues it also enables powerful actors — politicians, executives, and autocrats — to disclaim accurate video and audio as fake, leaving voters in a "nihilist bind where they do not know if they can trust anything."
Evidence and examples
Reported instances include Donald Trump's 2024 campaign, which reportedly called authentic Kamala Harris rally footage AI-generated, and authoritarian regimes that routinely claim documented atrocities are "deepfakes." Platform pre-bunking research cited by Persily indicates that awareness of deepfakes correlates with increased skepticism of authentic content rather than decreased skepticism.
Relation to policy
Persily frames the response around the principle that "we will not be able to 'tech our way out' of these AI-democracy challenges." He argues that watermarking and provenance authentication matter but are insufficient: content classification creates an expectation of verification, and once verification is expected, the absence of a label becomes a source of suspicion; mislabeling inevitably generates contestation that undermines the authentication system's credibility; and the emotional and psychological dimensions of mistrust, rather than the technical authentication gap, are the underlying problem.
Among measures Persily presents as helpful are provenance systems such as C2PA, described as necessary infrastructure that is not sufficient (Synthetic Content: Exploring the Risks, Technical Approaches, and Regulatory Responses); content-specific criminal statutes targeting the most severe harms such as non-consensual intimate imagery (TAKE IT DOWN Act — Source Summary); public-interest journalism that retains trust capital; and civil-society auditing combined with government-provided public compute for research access.
Relationships
- depends-on: Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security (Chesney & Citron, California Law Review, 2019) — origin of the term (Chesney & Citron, 107 Calif. L. Rev. 1753 (2019), section II.B.2.i).
- related: Robert Chesney, Danielle Citron — authors of the originating article.
- supports: The Digitalist Papers (Stanford, Volumes 1–2) — Persily essay treats this as the central democracy-AI threat.
- supports: Synthetic Media / Deepfakes / Synthetic Content: Exploring the Risks, Technical Approaches, and Regulatory Responses — technical mitigation landscape.
- supports: AI in Elections and Democratic Institutions — election-integrity framing.
- related: TAKE IT DOWN Act — Source Summary — legislative response targeting NCII subset of synthetic harm.
- related: The State of State AI: Legislative Approaches to AI in 2025 — state deepfake statutes.
- related: AI Snake Oil — Narayanan and Kapoor (2024) — Narayanan/Kapoor make a parallel empirical argument against deepfake-panic.
- contradicts: existential-deepfake-threat framing (common in 2023–2024 press coverage) — Persily argues this overstates the persuasion threat while underselling the trust-erosion one.