The video the candidate never filmed
On screen, a Senate candidate from Texas looks into the camera, standing in front of the state flag, in a blazer and a professional setting, and speaks for over a minute. The problem is he never filmed that video. The clip is an AI-generated ad from the National Republican Senatorial Committee (NRSC), which used deepfake technology to put James Talarico’s old social media posts — the Democratic nominee’s — into his mouth, on screen, in his likeness. The words “AI generated” appear in an easy-to-miss font in the lower right corner.
The detail that makes it novel is not that a fake video exists, but its length and its origin. It is the first to show a fake version of a candidate speaking in a lifelike manner for so long: an example of how far AI technology has come in a short time and an indicator of the direction attack ads may take. An 85-second deepfake, produced and distributed not by an anonymous troll but by a party’s official campaign apparatus, marks a different threshold from the forgeries that circulated before.
The ad’s technique is what makes it especially hard to fight. The Talarico ad mixed quotes from real tweets with completely fabricated commentary, creating a hybrid attack that even a forensics expert found nearly impossible to detect. It is not a pure lie, which would be easier to debunk, but a mix of authentic and synthetic material: the candidate did write those tweets years ago, but the digital version of him recites them and adds self-incriminating remarks he never made. The partial truth is the best wrapping for the forgery.
That hybrid mechanism is deliberate and especially effective against fact-checking. A debunk works when it can point out that something is false from start to finish; but when the material mixes authentic tweets with fabricated commentary, the fact-checker is trapped in a long, awkward explanation: it has to concede that part is real before clarifying which part is not. That initial concession dilutes the force of the debunk, because the viewer retains the “yes, he said that” and forgets the later nuance. The hybrid deepfake does not seek to deceive completely, but to sow just enough doubt that the distinction between what was said and what was fabricated becomes irrelevant in the voter’s memory.
From exception to strategy
What distinguishes 2026 from earlier cycles is not the existence of the deepfake, but its normalization as an accepted campaign tool. The Talarico ad is one of three recent ads from national Republicans that use deepfake technology. It is not an isolated slip, but part of a repertoire that is becoming routine.
The other cases confirm the pattern and show it crosses party lines. Senator John Cornyn circulated an AI-generated music video showing his Republican primary rival, Ken Paxton, committing corruption and infidelity, to a parody melody of the B-52s’ “Love Shack”; in January, Cornyn also posted a deepfake of Representative Wesley Hunt with no AI disclosure label. Among Democrats, the most notable user of AI-generated videos is California Governor Gavin Newsom. The tool belongs to no single side: both use it, which complicates any attempt to regulate it, because both parties benefit from the same techniques.
The defense offered by those who circulate them reveals the logic that normalizes the practice. The NRSC communications director defended the ad by saying Democrats were “panicking after seeing and hearing James Talarico’s own words.” The campaign of a Republican representative that used another deepfake defended it as “satire” and called it “the future of digital campaigning.” The argument is twofold: that the content is based on something real and that it is a legitimate form of political expression. Under that logic, the line between satire, attack and deception blurs conveniently.
The legal void that allows it
All this activity occurs on nearly ruleless terrain, and that void is deliberately hard to fill. There is no federal regulation constraining the use of AI in political messaging, leaving only a patchwork of largely untested state laws. In the absence of a national norm, each state legislates on its own, and the result is uneven.
The detail of those state laws shows why they offer little protection. Twenty-eight states — thirty by more recent trackers — have passed legislation on the use of AI in political ads, but most focus on disclosure rather than prohibition. That is: the typical law does not ban the deepfake, it only requires it to be labeled as such. And that approach has a flaw at its root. A label in small text in the corner of a video does not prevent the video from being clipped and reshared without context, nor a viewer from stopping after the first 30 seconds; the image sticks in memory after the disclaimer has scrolled past.
The laws that do prohibit have time windows the ads easily dodge. Texas law restricts political deepfakes within 30 days of an election, but the Talarico ad was circulated months before the November election and weeks before the May Republican primary runoff, outside that window. A restriction that applies only in the last month leaves the previous eleven months free, just when a campaign’s narrative is built. The rule exists, but its design makes it nearly inapplicable to real use.
The Cornyn-Paxton case: the internal war is also fought with AI
A detail of the US case deserves attention because it dismantles the idea that the deepfake is only a weapon between rival parties. The technology is also used inside, among allies competing for the same nomination. Senator John Cornyn, in his fight for the Republican primary, circulated an AI-generated music video portraying his rival Ken Paxton committing corruption and infidelity. It is an attack within his own party, not against the ideological adversary, which shows the tool serves any objective of power.
The case also reveals the inconsistency with which labels are applied. In January, Cornyn posted a deepfake of Representative Wesley Hunt with no AI disclosure label. When the same figure circulates a deepfake with a label in one case and without one in another, it is clear that compliance with the disclosure rule depends on the sender’s will, not on a mechanism that guarantees it. The rule that exists is followed at discretion, which amounts to having no rule in the cases that matter most.
The precedent came from far back and from high up. Already in the 2024 cycle, presidential campaigns such as the Florida governor’s used AI-generated images to persuade voters; the Republican national committee circulated in April of that year an entirely AI-generated ad showing a dystopian future, and the DeSantis campaign shared an attack with fabricated images of its rival. What in 2024 was a striking novelty, in 2026 became standard procedure. The adoption curve was two years: from experimental exception to campaign routine.
The regulation that never quite arrives
The US federal level has spent years trying, without success, to put rules on this practice. The Federal Election Commission advanced a petition to regulate ads that use AI to misrepresent political opponents as saying or doing something they did not, but the process has moved slowly. The difficulty is not only technical, but about incentives: when both parties benefit from the same tool, neither has real urgency to ban it.
That paralysis has a structural root the piece should name precisely. Congress has struggled for years to pass any meaningful technology legislation, and the political incentives to regulate campaign advertising are complicated when both parties can benefit from the same tools. The result is a void that will not close soon: the technology that enables the deepfake arrived before anyone was ready to govern it, and those who should govern it are the same ones who use it.
Some jurisdictions tried to get ahead with concrete penalties, but their reach is limited. A 2019 state law made it a misdemeanor to distribute deceptive deepfake content within 30 days of an election, with a penalty of up to a year in jail, but the NRSC ad was circulated months before the November midterms and weeks before the May primary runoff. The pattern repeats state after state: prohibitions exist, but confined to windows so narrow that most of the campaign falls outside their reach. The law punishes the last-minute deepfake, not the one that builds a narrative over months.
Why the label is not enough
The consensus among experts is that the dominant solution — requiring an “AI-generated” label — does not solve the underlying problem. The reason is how attention actually works on social media. A person scrolling fast through their feed does not read the small text in a corner; they see a familiar face saying something and move on. A 2025 peer-reviewed study found that people struggle to identify deepfake videos and that their political opinions are measurably affected by this type of misinformation. The harm does not require the viewer to believe forever; the momentary impression is enough.
The ecosystem that should slow the spread has weakened, not strengthened. Platforms like Meta and X label some AI-generated political content, but both dismantled their professional fact-checking programs and replaced them with community notes that move far slower than viral content. Speed is the decisive factor: a deepfake can reach millions before any correction mechanism — human or community — catches up with it. The correction always arrives late, once the impression has formed.
The cumulative effect worries more than any individual ad. A Purdue University professor who has studied thousands of deepfakes warned that the growing use of political content that spreads misinformation risks further eroding US voters’ trust in institutions. The risk is not only that a video deceives about a candidate, but that the mere existence of deepfakes makes voters distrust everything, including authentic material. When nothing is verifiable, even the truth becomes suspect.
The “liar’s dividend”
That last point has a name among disinformation scholars: the liar’s dividend. It is the idea that, once the public knows deepfakes exist and are convincing, any public figure can dismiss an authentic, compromising video by calling it fake. The technology that allows fabricating lies also offers an alibi to deny truths. In an environment saturated with forgeries, doubt becomes the default state, and that doubt benefits whoever wants to evade accountability.
Public perception already reflects that anticipated erosion. Research shows that 58 percent of US adults expect synthetic lies to escalate before votes are cast. When nearly six in ten voters expect to be deceived, the damage to trust is already done, regardless of how many deepfakes ultimately circulate. The expectation of deception is, in itself, a form of democratic harm.
It is best, however, not to fall into absolute catastrophism, because the evidence is nuanced. In 2024, AI’s actual impacts were far less than initially feared, and the public proved resilient against the deepfake threat by remaining skeptical of online information. People’s capacity to doubt — the same that feeds the liar’s dividend — also works as a defense against direct deception. Skepticism is at once the problem and part of the protection.
What it teaches the democracies that come later
For Latin American countries with elections ahead, the 2026 US case is a preview and a warning. The first lesson is that regulation always arrives behind the technology: by the time a law defines which deepfake to ban, the tool has already changed. Betting everything on legal prohibition is chasing a moving target. The second is that mandatory disclosure — the label — is necessary but insufficient, because it does not survive the clip nor the speed of the feed.
The third lesson, more hopeful, is where the defense does work. Experts’ recommendations point to content provenance: that campaigns and media mark their authentic material with verifiable credentials, which serves two functions at once: it gives a trust signal to voters and prevents the risk of being deepfaked. Instead of chasing every forgery, the strategy is to certify the true: to build a verifiable source of authentic material against which to contrast the imitations. It is a defense a country can mount without waiting for a federal law.
The fourth, and perhaps most important, falls on the voter. Fact-checking networks distribute reliable news to educate voters about deepfakes and the wider world of disinformation, especially in Spanish-speaking communities. Media literacy — teaching people to doubt what they see and to verify before sharing — is the slowest defense to build but the most resistant, because it does not depend on a platform or a government acting in time. In a world where forgery is cheap and correction is slow, the citizen’s informed skepticism is the last line.
The balance
The 2026 US midterm elections confirmed that the deepfake crossed from the shadows into the light: it is no longer a clandestine weapon, but an ad signed by official campaign committees of both parties. The technology advanced faster than the law, the platforms weakened their defenses just when they were most needed, and the dominant disclosure tool — the label — proved insufficient against speed and clipping.
The verdict is not that democracy is doomed by fake videos, because the 2024 evidence showed the voter retains significant resilience. The verdict is subtler and more unsettling: the greatest harm is caused not by the deepfake that deceives, but by the atmosphere of generalized doubt its existence installs, where the voter no longer knows what to believe and the liar finds an alibi to deny the true. The United States arrives at November having normalized a tool its own institutions do not yet know how to govern. For democracies holding elections later, the episode is a manual of what is coming and, above all, of what is worth preparing before the first fake video of a local candidate appears on screen.