Synthetic Endorsements and the Death of Political Trust

The endorsement is one of the oldest mechanisms in democratic politics. Someone you trust recommends a candidate. You extend some portion of your trust in that person toward the candidate they’re recommending. The chain of trust scales outward from individual relationships — your neighbor’s yard sign, your pastor’s Sunday mention, your union’s official position — until it aggregates into social proof at the community level. It is a distributed authentication system, and it has worked, roughly, for as long as representative democracy has existed.

What happens to that system when anyone can fabricate the authentication signal?

This is not a hypothetical. In the 2025 Philippine midterm elections, deepfake videos of former President Rodrigo Duterte — who by then faced ICC prosecution and had maintained public silence — circulated widely, showing him endorsing specific senatorial candidates. The videos were viewed an estimated 12 million times before Meta and TikTok completed their review processes. Two of the candidates shown receiving the synthetic Duterte endorsement won their races. Whether the videos affected the outcome is unknowable. What is certain: millions of Filipino voters received false information about who Duterte supported, believed it to varying degrees, and cast votes in a context where that false information existed.

The Trust Chain Has One Vulnerability

The genius of the endorsement as a political institution is that it is, by nature, public and attributable. If a union endorses a candidate, the union announces the endorsement under its name, through official channels, with its reputation attached. If the endorsement turns out to have been made under pressure or against members’ wishes, the union pays a reputational cost. The attributability creates accountability, and the accountability creates a rough incentive for honest signaling.

Synthetic endorsements break this mechanism at its foundation. The endorser did not endorse. The endorsement is detached from any actual human decision. There is no reputation at stake for the endorser because the endorser made no choice. The content exists, it circulates, it shapes impressions — but the accountability structure that gave endorsements their credibility is simply absent.

Related reading

The harm compounds because the response to a false endorsement is structurally disadvantaged relative to the false endorsement itself. When the Wisconsin business owner discovered the synthetic endorsement video in November 2025, his options were: issue a public denial (which requires that his denial reaches the same people who saw the false video — it almost never does), pursue legal action (which takes years and produces no pre-election remedy), or ask platforms to remove the content (which some did, eventually, after varying delays). The original content had already done its work.

This asymmetry — cheap to create, expensive to correct — is the defining feature of the synthetic endorsement problem, and it is not an engineering problem. It is a structural feature of how information propagates in attention economies.

The Categories of Harm

Synthetic political endorsements harm democratic processes in at least three distinct ways, and conflating them makes it harder to think about responses.

The first is the direct falsification of specific endorsements. A real person appears to support a candidate they do not support. The harm is specific and attributable: votes are cast based on false information about where a trusted figure stands. This is the Wisconsin case, the Philippine case, and hundreds of smaller incidents across state and local elections in 2025 and 2026.

The second harm is subtler and probably more significant: the general degradation of the endorsement as a credible signal. When synthetic endorsements become common enough, voters rationally discount all endorsements. The business owner endorsing a candidate on video — how do you know that’s real? The union’s official statement — was that actually decided by the membership, or was the video of the general secretary reading it manufactured? The skepticism is not irrational. It is the correct epistemic response to an environment where fakes are common and indistinguishable from originals. But the logical response to individual false signals is a corrosive uncertainty that undermines genuine social proof.

The third harm is what political scientists call the “liar’s dividend” — the ability of bad actors to deny genuine statements by claiming they are synthetic. A candidate caught on video making a damaging statement in a private setting can, in 2026, claim the video is AI-generated and face an information environment where that claim is taken seriously, because synthetic videos do exist, and because the technical capability to verify authenticity is not widely distributed. The liar’s dividend is the mirror image of the false endorsement problem: instead of false content being believed, true content becomes deniable.

Local Politics, Fewer Resources

National political races — presidential campaigns, high-profile Senate contests — have fact-checking operations, legal teams, and communications infrastructure that can, imperfectly and slowly, respond to synthetic content. Smaller races do not.

The 2026 election cycle has seen the most damaging synthetic endorsement content deployed at the state legislature, city council, and school board levels. These are races where the total budget might be $50,000. The candidate is running on a few evenings per week while holding a day job. The local newspaper may no longer exist. The media ecosystem that might catch and correct false information is attenuated or absent.

A synthetic video showing a school board candidate making racist statements at a private gathering — the candidate never attended any such gathering, the video was entirely generated — circulated in a suburban district outside Columbus, Ohio in February 2026. The candidate, Maria Chen, a second-generation Chinese-American who had run on an inclusive curriculum platform, lost by 847 votes. The video had been viewed by approximately 22,000 people in a district with roughly 18,000 registered voters. It was confirmed as synthetic nine days after the election. No remedy was available.

Chen’s case is not exceptional. It is representative of what happens when AI-generated political deception encounters the thin information environment of local politics. National political events have dense enough media coverage that synthetic content faces more scrutiny. Local politics is almost entirely unsupervised, and local decisions — school boards, city councils, county commissions — are where most of the governance most people experience actually happens.

Related reading

The Trust Market Has Collapsed Before

There is a useful historical analogy in the collapse of product trust during the patent medicine era of the late nineteenth century. Before the Pure Food and Drug Act of 1906, the market for medical products was essentially unregulated, and it filled with fraudulent testimonials. Famous physicians (sometimes real, sometimes invented) endorsed products that contained opium, cocaine, or simply flavored water. Endorsements meant nothing because any endorser could be fabricated and any claim could be made. The result was not that people stopped buying patent medicines — they kept buying them, desperately, because they had real health needs and limited alternatives. The result was that they couldn’t rationally distinguish good products from bad ones, and the entire market for credible health information collapsed into noise.

The political endorsement market in 2026 is beginning to look like the patent medicine market in 1900. The signal is corrupted. Sophisticated voters respond by increasing their skepticism, which is epistemically correct but practically limiting — skepticism doesn’t help you decide whom to vote for. Unsophisticated voters (most voters, including most of the sophisticated ones in domains they haven’t personally analyzed) respond by continuing to be influenced by the signals they receive, because the alternative is trying to verify everything from first principles, which is cognitively impossible at election scale.

The Pure Food and Drug Act worked because it created a third-party verification institution — government standards and labeling requirements — that substituted for the collapsed private authenticity market. It didn’t require consumers to personally verify every claim. It required producers to submit to testing, and it punished fraudulent claims with legal consequences that acted as deterrents.

The equivalent for synthetic political content would be a third-party verification architecture for political media — cryptographic authentication for political communications, provenance tracking built into video and audio platforms, legal penalties that create deterrence for fabrication. These things are technically feasible. They are politically difficult, because the politicians who would have to pass the relevant legislation are the same people who benefit from the ambiguity.

The Adversarial Dynamic

Any defensive measure against synthetic endorsements faces an adversarial dynamic that cryptographic authentication alone cannot solve.

Detection tools improve. Generation tools improve faster. In January 2026, the most sophisticated detection software could identify AI-generated video with approximately 89 percent accuracy on fresh content and about 63 percent accuracy on content that had been compressed and re-encoded through social media platforms (which is how essentially all political content actually travels). By March 2026, a new generation of generation tools had reduced the detection rate by another 15 percentage points. The attacker-defender race has consistently favored the attacker, because generating plausible synthetic content is computationally easier than definitively identifying synthetic content.

Cryptographic authentication — embedding provenance information into video at the time of creation, signed by the camera hardware — is the more promising approach. The Content Authenticity Initiative, backed by Adobe, the BBC, the New York Times, and others, has made progress on implementing C2PA standards (Coalition for Content Provenance and Authenticity) for media. If every authentic video carries a verifiable signature from the device that captured it, the absence of such a signature becomes evidence of fabrication.

The problem is adoption. C2PA works if the cameras that capture political events implement the standard. Major camera manufacturers have begun adding hardware signing to new devices. Most of the video in circulation is not captured on new, premium devices with hardware signing — it’s captured on older smartphones that don’t implement the standard, or captured legitimately on new devices but then re-encoded in a way that strips the provenance metadata. The standard exists. The ecosystem for making it meaningful does not yet.

What Endorsements Were Actually For

There’s something worth preserving here that usually gets lost in the technical discussion of detection and authentication.

Political endorsements exist because democratic decision-making at scale is cognitively impossible without heuristics. An average voter in a November election is choosing between candidates for twenty to forty offices, many of which she has not had time to research independently. Endorsements — from unions, newspapers, civic organizations, trusted community figures — are heuristics that allow voters to make reasonably informed decisions in constrained time. They are, in a genuine sense, part of the infrastructure of democratic participation.

Related reading

When that infrastructure is corrupted, the harm is not just to individual candidates who receive false endorsements. The harm is to the distributed capacity of a democratic society to aggregate genuine preferences. Elections become less accurate measures of what people actually want when the information environment surrounding them is systematically unreliable.

This is the argument that goes beyond the individual injustice to Maria Chen in Columbus or the business owner in Wisconsin. The individual injustices are real. But they are symptoms of a more fundamental problem: an information environment in which the social mechanisms through which democratic societies form and communicate political preferences have been compromised by tools that are cheap to use, impossible to reliably detect, and largely unregulated.

Fixing that environment requires more than better detection software. It requires deciding, as a matter of democratic self-preservation, what the cost of political deception should be — and making that cost high enough to change the behavior of people who currently face essentially no consequences for fabricating political content.

That is a political problem. It is unsolved.