It is late in an election cycle. Polls close in forty-eight hours. A video appears online — at first dismissed as another fragment in the daily churn of political media. Then it sharpens. A candidate sits across from a donor. Money changes hands. A sentence is spoken that appears to confirm the worst suspicions of opponents and the deepest fears of supporters. Within minutes, the clip has spread across platforms. Within an hour, it has been viewed millions of times. Campaigns scramble to respond. Journalists begin to investigate. Fact-checkers move as quickly as they can.
By the next morning, experts conclude the video is a fabrication — generated by artificial intelligence. But the damage is done. Early ballots have already been cast. Undecided voters have already incorporated doubt into their judgments. The election proceeds on schedule. Reality, however, has lagged behind.
This scenario is not hypothetical in the way it would have been even two years ago. The 2026 midterm cycle is proving to be the first major American election in which deepfake political advertising is being deployed at scale, by major party committees, against named candidates, in competitive races — and the institutional apparatus that is supposed to prevent it from working is operating at a speed that makes prevention structurally impossible.
around recent U.S. primaries
(no federal regulation exists)
anti-deepfake election task force
Truth operates on investigative time. Democracy operates on electoral time. Deepfakes operate on viral time. Only one of these moves at the speed of the modern information environment — and it is not the one democracy depends on.
The Fabrication Is Already Here
The first confirmed deepfake advertisement in a United States Senate race appeared this month. The National Republican Senatorial Committee produced an AI-generated video of Graham Platner, a Democratic Senate candidate in Maine, showing a fabricated shirtless version of the candidate reading controversial Reddit posts he had written ten to fifteen years ago. The words "AI generated" appeared in small text in the lower right corner. The video was realistic enough that Platner's campaign issued a public statement warning about the weaponization of AI in political advertising.
It was not an isolated case. In Georgia, the campaign of Republican Representative Mike Collins, who is challenging Democratic Senator Jon Ossoff, produced a deepfake video in which Ossoff appears to say: "I just voted to keep the government shut down. They say it would hurt farmers, but I wouldn't know. I've only seen a farm on Instagram." Ossoff never said this. The Collins campaign spokesperson told Reuters that as technology evolves, the campaign "will be at the forefront embracing new tactics and strategies." In California, Governor Gavin Newsom — a potential 2028 presidential candidate — has been using deepfake videos to troll President Trump, establishing a pattern of AI-generated political content that crosses party lines.
These are not underground operations. They are official campaign products, produced or commissioned by national party committees and named candidates, distributed through standard advertising channels and social media, labeled in fine print that most viewers will never notice, and deployed in an environment with no federal regulation constraining their use. Twenty-eight states have passed legislation addressing AI in political ads, but most focus on disclosure requirements rather than prohibitions, and none has been meaningfully tested in enforcement. Maine's governor signed a disclosure law last month. It has not yet taken effect.
A 2025 study published in the peer-reviewed Journal of Creative Communications found that people struggle to identify deepfake videos and that their opinions are measurably affected by this kind of fabricated content. Daniel Schiff, a Purdue University professor who has studied thousands of deepfakes, told Reuters that "the types of damage that we can do to the rigor and credibility of elections and democratic systems very much risks being supercharged." Identity-verification firm Sumsub reported deepfake incidents rising 303% around recent U.S. primaries. The World Economic Forum ranked AI-driven misinformation as the single greatest short-term global risk in both its 2025 and 2026 Global Risks Reports.
The Three Clocks
The temporal mismatch in deepfake politics operates across three distinct timescales, and the misalignment between them is the structural problem that no current institutional response has solved.
The fabrication clock runs at viral time. The production of a convincing deepfake video now requires minutes rather than days. Widely available generative AI tools allow individuals with minimal technical expertise to produce audio and video that is difficult for untrained viewers to distinguish from authentic recordings. The distribution of such content is even faster — falsehoods move through social networks at a speed that outpaces correction, often reaching millions of viewers before verification has even begun. A deepfake posted at 9:00 p.m. can be viewed ten million times before the first authoritative fact-check is published at noon the following day. The production-to-saturation cycle now operates on a timescale of hours.
The verification clock runs at investigative time. Fact-checking, even when accelerated by new tools and coordinated responses, still requires time. Careful analysis, source validation, forensic examination, expert consultation, and public communication cannot be compressed indefinitely without sacrificing the credibility that makes fact-checking valuable in the first place. A rushed fact-check that turns out to be wrong — declaring a real video fake, or vice versa — does more damage than no fact-check at all. The verification cycle operates on a timescale of hours to days, and the quality of the output is inversely proportional to the speed at which it is produced.
The democratic clock runs at electoral time. Elections are fixed points. Voters must act on the information available to them at the moment of decision — not on the information that will become available after the ballots are counted. Early voting, which now accounts for a substantial fraction of ballots cast in many states, extends the window during which voters are making decisions on incomplete or contested information. A deepfake that appears three days before an election and is debunked two days later has already influenced every early ballot cast during those forty-eight hours. The democratic clock does not wait for the verification clock to finish its work.
The regulatory clock runs slowest of all. Legislative processes unfold over months or years, constrained by political negotiation, legal scrutiny, constitutional challenges, and the procedural tempo of democratic lawmaking. The twenty-eight state laws addressing AI in political ads were enacted over a period of years. No federal legislation exists. The gap between the speed at which deepfakes can be produced and the speed at which the legal framework can evolve to address them is not closing. It is widening with every improvement in generative AI capability.
The fabrication of a convincing falsehood now takes minutes. The verification of that falsehood takes hours to days. The regulatory response to the category of threat takes months to years. The election happens on a fixed date that waits for none of these processes to complete. The asymmetry is structural: it cannot be resolved by making any single clock faster, because the clocks are operating on fundamentally different temporal architectures.
The Dividend
The concept that gives this dispatch its name comes from the deepfake research literature, and it describes a second-order effect that is more dangerous than the deepfakes themselves.
The liar's dividend is this: once the public understands that any piece of audio or video can be plausibly fabricated, the evidentiary value of all recordings diminishes — including real ones. A politician caught on camera saying something damaging can now claim the video is a deepfake, and a meaningful fraction of the audience will accept that claim, because the possibility of fabrication has been established as a permanent feature of the information environment. Real scandals can be reframed as digital fabrications. Genuine recordings can be cast into doubt. The deepfake does not need to be believed to do its work. It only needs to exist as a category — as a known possibility — to give every actor in the system plausible deniability about every piece of recorded evidence.
This is a qualitatively different problem from misinformation. Misinformation introduces false claims into the information environment. The liar's dividend removes the shared framework for determining what is true. The first is a problem of signal — too much noise in the channel. The second is a problem of epistemology — the channel itself can no longer be trusted.
For much of the modern era, recorded audio and video served as a final arbiter of disputed claims. "The tape" was the thing that settled arguments, ended careers, resolved contested accounts of what happened. That assumption no longer holds. When any piece of media can be plausibly fabricated, the concept of definitive evidence begins to dissolve. Verification becomes more complex, more time-consuming, and more uncertain. And voters — who must make decisions within fixed electoral windows — are not simply misinformed. They are asked to make decisions in the absence of a stable framework for determining what is real.
The consequences are not theoretical. They are already visible in the way political actors respond to unfavorable evidence. The reflexive claim "that's AI-generated" or "that's been manipulated" is now a standard component of crisis communications, deployed whether or not there is any reason to believe fabrication has occurred. The category of the deepfake has given every public figure a standing escape hatch from accountability, and the escape hatch works not because the audience is gullible but because the audience has correctly understood that the technology exists and that fabrication is possible. The liar's dividend is paid in the currency of reasonable doubt, and reasonable doubt is the one thing democratic institutions cannot afford to have attached to the evidence on which elections are decided.
The Korean Experiment
Against this backdrop, South Korea has undertaken one of the most ambitious efforts in the world to align institutional response speed with the velocity of the deepfake threat. What it has built is the closest thing to a counter-example that the temporal-mismatch thesis has encountered in nine dispatches — and its limitations are as instructive as its ambitions.
Amendments to South Korea's Public Official Election Act prohibit the production, distribution, or posting of AI-generated deepfake content for electoral purposes within ninety days of any election. Violations carry severe penalties, including prison sentences and substantial fines. The law does not rely on labeling requirements or partial restrictions. It imposes a categorical ban within a defined temporal window.
In advance of nationwide local elections scheduled for June 3, 2026, authorities have escalated enforcement to an extraordinary scale. A dedicated task force of 596 personnel — drawn from the prosecution service, the National Police Agency, and the National Election Commission — has been placed on standby. Five categories of electoral crime fall under a zero-tolerance mandate, with deepfake disinformation drawing the most attention both domestically and internationally. The task force is supported by technical infrastructure: government agencies have deployed AI-based detection systems capable of analyzing suspicious media in real time, with reported accuracy rates exceeding ninety percent. The Ministry of Science and ICT is funding deepfake detection and blocking technology. The prosecution service and police have been authorized to raid sites spreading fabricated content.
On its face, this represents a rare institutional attempt to match the speed of the threat. Detection is proactive rather than reactive. Enforcement is coordinated across agencies. Technical tools are integrated into the decision-making process. Penalties are severe enough to function as genuine deterrents. In contrast to the incrementalism that characterizes most democratic regulatory efforts, South Korea's model is notable for its scale, its clarity, and its willingness to act before the damage is done rather than after.
Yet the model also exposes the limits and tradeoffs inherent in speed-matching.
Enforcement in a global information environment is inherently difficult. Content hosted on foreign servers, distributed through encrypted channels, or amplified by actors beyond national jurisdiction may evade rapid suppression. Even when detection is swift, removal does not erase initial exposure — impressions formed in the first hours of dissemination persist despite subsequent correction. Detection accuracy of ninety percent sounds impressive until you calculate what ten percent of a sufficiently large volume means: thousands of undetected fabrications circulating during the critical pre-election window.
And a system capable of rapidly identifying and suppressing certain categories of content raises questions that democracies cannot avoid. The capacity to act quickly is also the capacity to act broadly. The same mechanisms that prevent harm can, under different political conditions, be used to constrain legitimate speech. The line between disinformation and protected expression is not always clear, particularly in politically charged contexts. South Korea's approach may depend on specific cultural and institutional conditions — a compact media environment, strong state capacity, established norms of regulatory enforcement — that are not easily replicated in larger, more fragmented democracies.
South Korea demonstrates that institutions can, under certain conditions, mobilize resources, authority, and technology to confront the deepfake threat with speed approaching the threat's own velocity. It also demonstrates how demanding such adaptation is — requiring political will, legal flexibility, technical capacity, cultural context, and sustained attention at a level that most democracies have not shown themselves capable of sustaining. If this is the leading edge of institutional response, the question it raises is uncomfortable: how many systems are capable of implementing comparable measures, and at what cost to other values?
The Arms Race
There is a version of the optimistic case that holds that detection technology will eventually close the gap — that watermarking, provenance tracking, forensic analysis, and AI-based detection will mature to the point where deepfakes can be identified reliably and quickly enough to prevent their effects.
The data does not support this version of optimism. The arms race between generation and detection is currently being won by generation, and the structural reasons for this are unlikely to reverse.
Deepfake generation tools are improving faster than detection tools because the incentive structures favor generation. Generative AI is a multi-billion-dollar industry with commercial applications far beyond political manipulation — entertainment, advertising, content creation, personalization — and the same improvements that make generative tools commercially valuable also make them more capable of producing convincing fabrications. Detection tools, by contrast, are a niche application with limited commercial value, funded primarily by governments and nonprofits, and their development cannot match the pace or investment of the commercial generative sector.
Watermarking and provenance systems like Content Credentials and Google's SynthID represent real progress, but they are fragile. Adversaries can migrate to unmarked open-source models that do not implement watermarking protocols. Detection tools that rely on identifying artifacts of AI generation become less effective as generation quality improves. Each new generation of AI models produces output that is harder to distinguish from authentic content, and the improvement cycle is measured in months rather than years.
Europol has projected that up to ninety percent of online content could be AI-generated by 2026 — a figure that, even if overstated, indicates the scale of the challenge. The platforms that host and distribute this content have moved in the opposite direction from robust enforcement: Meta and X have both scrapped professional fact-checking systems in favor of user-generated community notes, reducing the institutional capacity for content verification at the same moment the volume of content requiring verification is exploding.
The trajectory is not reassuring. Faster tools, more convincing outputs, and more sophisticated distribution networks will increase both the frequency and the impact of epistemic disruptions. The fundamental constraint remains: elections occur on fixed timelines, while truth requires time to establish. Detection and verification efforts may mitigate some effects, but they are unlikely to eliminate the underlying asymmetry between the speed at which falsehoods can be produced and the speed at which truth can be confirmed.
The Epistemic Foundation
Pull back from the specific mechanisms and the structural problem becomes visible in its full form.
Democracy presumes a degree of epistemic common ground. Not agreement — democracies are built to manage disagreement — but a shared set of reference points from which disagreement proceeds. Voters can disagree about what a policy means, what a candidate's record implies, what the right course of action is. But those disagreements are grounded in shared facts: shared recordings, shared documents, shared events that both sides accept as having occurred, even if they interpret them differently.
The deepfake problem threatens this foundation not by introducing false facts — propaganda has always done that — but by dissolving the category of the fact itself. When any recording can be plausibly fabricated, and when the verification of any recording takes longer than the decision window in which voters must act, the shared reference point ceases to function. Disagreements are no longer about interpretation. They are about whether the evidence exists at all. That is a different kind of problem, and it is one for which democratic institutions have no established remedy.
Elections continue to occur on fixed schedules. Ballots are cast. Outcomes are declared. The procedural aspects of democracy persist. But the legitimacy of those outcomes depends on a public willing to accept that decisions were made in reference to a common reality. If that condition no longer holds — if a sufficient fraction of the electorate believes that the evidence on which the other fraction based its decision was fabricated, or if both fractions are uncertain about the status of the evidence that informed their own decisions — then the procedural shell of democracy contains an epistemic void at its center.
This is not a prediction of democratic collapse. Democracies are resilient institutions with deep reserves of legitimacy, procedural inertia, and civic habit. But it is a description of the condition under which that resilience will be tested, and the testing has already begun. The 2026 midterms are the first major American election cycle in which deepfakes are being deployed as standard campaign tools by official party committees. Whatever happens in November will be the first data point in what is likely to be a long and difficult reckoning with the question of whether systems built on shared reality can endure when reality can no longer be reliably shared.
What This Changes
This dispatch extends the temporal-mismatch thesis into a domain that is, in one specific sense, more fundamental than any of the eight that preceded it.
The previous dispatches described institutions that were too slow to respond to external threats, or — in the case of the Immune Response — institutions being deliberately dismantled. In each case, the institution's failure affected a specific domain: military posture, financial stability, alliance cohesion, public health capacity, the production of state communication. The damage was real but bounded. The system being strained was a system within the larger democratic order.
The deepfake problem is different because it threatens the substrate on which all the other systems depend. Democratic military policy depends on an informed electorate. Democratic financial regulation depends on elected officials who were chosen on the basis of shared facts. Democratic public health policy depends on public trust in evidence. Democratic alliance management depends on voters who can evaluate their government's decisions against a stable factual record. Every institutional function that the previous dispatches examined presupposes that the democratic process producing the leaders and policies governing those institutions is itself epistemically sound — that voters are making decisions in reference to a reality they can, with effort, verify.
The liar's dividend attacks that presupposition. It does not attack any single institution. It attacks the epistemic foundation on which all democratic institutions rest. If voters cannot agree on what is real, the quality of every downstream decision — military, financial, diplomatic, public health, technological — degrades, because those decisions are being made by leaders chosen through a process whose informational integrity has been compromised.
Democracies were designed in an era when reality could be verified before decisions had to be made. That assumption is increasingly untenable. The challenge is no longer confined to regulating content or improving detection. It concerns the conditions under which collective decisions are possible at all. If reality itself becomes persistently contestable within the timeframes that elections demand, then the problem is not merely technological or legal. It is structural. And the question facing modern democracies is whether systems built on shared reality can endure when reality can no longer be reliably shared.
Truth operates on investigative time. Democracy operates on electoral time. Deepfakes operate on viral time. The gap between them is not closing. It is the gap in which the legitimacy of self-governance is being decided.
◆ End of dispatch
Nine dispatches. Nine domains. One pattern — now visible in its most complete form.
The One-Year Threshold showed the temporal mismatch in military endurance. The Ugarit Warning showed it in civilizational stress. The Signature Problem showed it in cryptographic governance. The Abdication Instinct showed it in machine alignment. The Toll Booth and the Tripwire showed it in alliance politics. The 280-Character Cabinet showed it in state communication. The Amplification Channels showed it in global finance. The Immune Response showed it in pandemic preparedness — and introduced the failure mode of deliberate institutional dismantling.
The Liar's Dividend shows the temporal mismatch in the deepest layer of the system: the epistemic foundation on which democratic decision-making depends. Every previous dispatch examined a specific institutional domain that was failing to keep pace with a specific threat. This dispatch argues that the substrate beneath all those domains — the shared factual reality that democratic governance presupposes — is itself being eroded at a speed that the institutions designed to protect it cannot match.
Nine domains. One pattern. Three variations of the pattern. Institutions that are too slow (dispatches one through seven). Institutions that are being deliberately dismantled (dispatch eight). And now: the epistemic foundation on which all institutions depend being eroded faster than it can be repaired (dispatch nine).
The series has one dispatch remaining. It will bring the thesis home — not to another system, but to the reader. The temporal mismatch is not only a property of institutions. It is a property of individual lives, and the gap between the speed at which the world is being reorganized and the speed at which any person can adapt to the reorganization is where the most personal version of the pattern lives.
Complex systems fail not when the threat arrives but when the threat arrives faster than the institutions can process it. The deepest version of that failure is when the institutions can no longer establish what is real before the people they serve must decide what to do.