The fear that dominated the 2020 and 2022 cycles — that deepfake video would be deployed in a major presidential race to put fabricated words in a candidate’s mouth — has not materialized in the way predicted. No major deepfake video of a presidential or Senate candidate has achieved the kind of mainstream viral moment that the threat analysis envisioned. This has been taken in some quarters as evidence that the deepfake threat was overstated.
It wasn’t. The threat materialized differently than predicted.
The sophisticated political use of AI-generated video has concentrated not at the top of the ballot, where media scrutiny is intense and campaigns have the resources to respond rapidly, but at the bottom — in state legislature races, city council elections, school board contests, primary elections, and the hundreds of races that decide real governance but operate in thin information environments with minimal oversight. These are the elections that have actually been affected. They don’t generate national news coverage, so the pattern they collectively represent has not been adequately recognized.
The Technology Capability in 2026
Let me be specific about what is technically possible, because the discourse often operates at a vague level that obscures the reality.
A video in which a real person’s face appears to say words they never said — what is commonly called a deepfake — can be produced in 2026 by any technically literate person with access to a standard laptop and publicly available software. The quality available to a non-expert with free tools is roughly equivalent to the quality that required a $50,000 professional production budget in 2022. A competent technical professional with commercial software and a few hours of work can produce video that is convincing enough to fool most viewers who are not specifically looking for signs of fabrication.
Voice cloning — replicating a specific person’s voice from a sample of their audio — is, if anything, more advanced than video synthesis. A convincing voice clone of a public figure with several hours of available audio can be produced in under a day from a sample of as little as three minutes of clean audio. The cloned voice can then say anything.
Related reading
The limiting factor in 2026 is no longer the quality of the fake. It is the availability of authentic comparison material for verification. Deepfake video of a politician who has appeared on television thousands of times, whose speaking patterns and mannerisms are well-documented, is easier to spot because the authentic reference corpus is large and the anomalies are detectable. Deepfake video of a first-time candidate who has given three public speeches and whose face is known only to a small local community is much harder to verify as fake, because there’s very little to compare it to.
The Cases at the Bottom of the Ballot
The pattern across the 2026 cycle is consistent enough to constitute a trend.
In a February 2026 Democratic primary for a state house seat in Nevada, a video circulated showing the leading candidate, a local teacher named Sandra Morales, making statements about charter schools that were deeply at odds with her actual positions. The video was shared in a parents’ Facebook group with 4,000 members. Morales lost the primary by 340 votes. The video was confirmed as synthetic by a forensic analysis commissioned three weeks after the election. No remedy is available; primary results are not subject to reversal for information environment violations.
In an April 2026 city council race in a mid-sized Virginia city, a synthetic audio recording circulated in a largely Black neighborhood, in the voice of the Black candidate’s primary white opponent, making racially inflammatory statements. The recording drove significant social media activity and coverage from one local blog. The white candidate lost. A subsequent investigation suggested the audio had been created by a third party with no clear campaign affiliation, possibly targeting the opponent as a form of “dirty trick” designed to benefit neither official candidate but to inflame racial tensions in the district.
In a May 2026 school board race in a suburban district in Texas, a video appeared to show a Democratic candidate expressing support for explicit sexual content in school libraries. The video was synthetic. It was shared approximately 14,000 times before the election. The candidate lost by 800 votes.
These are three cases. The actual number of state and local elections in the 2026 cycle in which AI-generated video or audio played a documented role is, as of this writing, approximately forty — according to the tracking maintained by the Brennan Center for Justice, which began monitoring the issue in January 2026. The actual number, including cases that have not been documented or investigated, is unknown but is believed by researchers to be substantially higher.
The Detection Problem Is Worse Than Reported
The public understanding of deepfake detection is about two years behind the current state of the technology, in both directions.
Most people believe that deepfakes are detectable by trained experts if examined carefully. This was true in 2020 and partially true in 2022. In 2026, it is true of some deepfakes and false of others. The top-tier generation systems, deployed by well-resourced actors, produce content that is genuinely difficult for forensic analysis to identify with high confidence. The degradation that occurs when content is compressed, re-encoded, and re-shared through social media further reduces detection accuracy.
The automated detection systems deployed by Meta, YouTube, and others detect AI-generated content with roughly 70 to 80 percent accuracy on fresh, high-quality content, and much lower accuracy on content that has been re-encoded. They apply these systems to political content with various levels of reliability — higher reliability for content flagged by users, lower reliability for content in low-review-rate languages and regions.
The practical implication: detection works as a probabilistic tool that catches some proportion of AI-generated political content, not as a reliable gatekeeper. A bad actor deploying deepfake video at scale, with some technical sophistication about how detection systems work, will generate enough content that some significant fraction reaches its intended audience before detection.
The reverse problem is also real. False positives — authentic video incorrectly identified as AI-generated — are a documented problem, and the error rate is not evenly distributed. Content in languages and dialects underrepresented in detection system training data has higher false-positive rates. Content featuring people with physical features that are less represented in training data is more likely to be incorrectly flagged. The detection systems’ errors are not random. They systematically disadvantage content produced by and for underrepresented communities.
Related reading
When the Fake Is the Attacker’s Cover
There is a deepfake use case that generates less coverage than false videos of opponents but is arguably more insidious: fabricated content deployed to discredit authentic damaging material.
In the summer of 2026, authentic audio recordings circulated purporting to show a congressional candidate in the Southwest making private statements about fundraising donors that were legally and ethically problematic. The candidate’s campaign responded by claiming the recordings were AI-generated — an assertion made without evidence but in a media environment where that claim is plausible and difficult to quickly disprove. The candidate commissioned a private forensic analysis that, its authors said, found “indicators consistent with AI generation.” Independent forensic analysts who reviewed the same material reached different conclusions. The ambiguity persisted through the election.
The candidate won.
Whether the recordings were authentic or synthetic is disputed as of this writing. What is not disputed: the claim that they were synthetic provided effective cover against authentic damaging information, because in an environment where deepfakes exist and are common, “that’s AI-generated” is a plausible defense that takes time to refute and may never be definitively refuted.
Political strategists have a name for this: the liar’s dividend. The existence of deepfakes benefits dishonest actors not only through the creation of false content but through the ability to deny authentic content. The more sophisticated and prevalent deepfakes become, the more valuable that denial capacity is.
This means the deepfake threat is not symmetrical between creators of false content and targets of false content. The threat actually benefits dishonest actors in two directions: they can create false content about opponents and they can deny authentic content about themselves. Honest actors who don’t create false content get only the downside of the technology — the risk of being victimized by synthetic content they didn’t create.
The Consent Question in the Age of Digital Likeness
There is a legal and ethical dimension to deepfakes that the political conversation often bypasses in its focus on disinformation effects: the question of consent to the use of one’s likeness.
When a political operative creates a video in which a real person appears to say things they never said, that person’s face and voice — their identity, in a meaningful sense — have been appropriated without consent. This is true regardless of the political context. It is true when the victim is a politician and it is true when the victim is a private citizen. The Wisconsin business owner and Maria Chen in Columbus were private individuals whose likenesses were used by political actors without consent.
The legal remedy for this is tort law — specifically, the right of publicity and defamation claims — which is slow, expensive, and produces no pre-election remedy. Some states have passed legislation specifically creating criminal liability for non-consensual deepfakes in political contexts. California, Texas, and Georgia have such laws. They are difficult to enforce against anonymous actors and have not produced significant deterrence in the 2026 cycle.
The deeper issue is philosophical. Democratic theory has generally held that candidates for public office accept a reduced expectation of privacy and accept aggressive political speech about their public roles. Criticism, satire, and even sharp attacks on candidates have constitutional protection. The question of whether synthetic impersonation — putting fabricated words and fabricated images in a real person’s identity-envelope — is protected political speech or prohibited identity theft is one that legal systems are only beginning to grapple with.
The answer to that question has implications beyond politics. The same technology that creates a fake political candidate video also creates non-consensual synthetic pornography, fraudulent identity impersonation in financial contexts, and fabricated evidence in legal proceedings. The legal status of synthetic identity impersonation is a foundational question about digital rights that the political deepfake problem dramatizes but doesn’t uniquely define.
Related reading
What Would Actually Help
Three interventions would meaningfully reduce the harm from deepfake political content, and they are ordered here not by effectiveness but by political feasibility.
The most feasible: civil liability reform that makes deepfake creators financially liable for documented electoral harms. Currently, a victim of a synthetic endorsement video has a defamation claim that takes years to resolve and requires proving specific damages — a difficult standard in electoral contexts. A reformed liability standard that created presumptive damages for AI-generated political content featuring a real person without their consent, with trebled damages for content deployed within ninety days of an election, would create financial deterrence without requiring proof of specific electoral harm. This kind of reform has support across partisan lines and doesn’t raise the same First Amendment concerns as content prohibitions.
Moderately feasible: mandatory cryptographic provenance for political advertising content. Campaigns and political organizations filing with the FEC or state equivalents should be required to submit political advertising content with cryptographic provenance chains — proof that the video was captured by an actual camera at a real event. This applies only to paid advertising, not to organic content, but it would raise the standard for the formal advertising environment while leaving organic content distribution to other regulatory approaches.
Least feasible but most impactful: platform-level accountability for demonstrably false political content that causes specific documented harm. This runs directly into First Amendment doctrine, strong tech industry lobbying, and the genuine difficulty of defining “demonstrably false” in political contexts. The political and legal barriers are formidable. The argument for it is simple: when a platform profits from distributing content that is fabricated and causes specific documented harm to specific individuals, the platform should bear some responsibility for that harm.
The technology will not wait for these reforms to be enacted. The question is whether reforms come before or after the harms become clearly irreversible.
One email a month: new articles, reviews and the upcoming live webinar + free recording. No spam, unsubscribe anytime.