Technically Disclosed - Decipon Insights

The Ad

In March 2026, the National Republican Senatorial Committee released an attack ad against James Talarico, the Democratic nominee for US Senate in Texas. The ad showed a man who looked and sounded exactly like Talarico delivering a direct-to-camera monologue about his past statements on transgender issues, race, and religion. The figure recited actual tweets Talarico had written years earlier. It also, according to the candidate himself, included invented statements he never made.

In the bottom right corner of the screen, for approximately three seconds near the beginning, the words “AI GENERATED” appeared in small text. Then, as the synthetic Talarico continued to speak, a fainter and smaller version of the same text remained in the same corner for the duration.

The NRSC called the ad “In His Own Words.”

The words were not his. The face and voice were not his. The disclosure was real. The ad ran.

The Disclosure That Isn’t

There is no federal law governing the use of AI in political advertising. There is no federal requirement that AI-generated content in campaign ads be disclosed at all. The NRSC created its own disclosure standard, met it, and published an attack ad in which a synthetic version of a real Senate candidate said things on camera that the candidate did not say.

The disclosure tells you something was made with AI. It does not slow the video, interrupt the audio, or interrupt the emotional experience of watching a person apparently speaking directly to you. Three seconds of small text in a corner is where lawyers live — it’s the same register as the “Paid for by” tag that scrolls past at the end of a television spot. It satisfies no viewer. It creates a paper trail.

This is the move: proactive compliance theater. Do the minimum possible that could, under a hypothetical future legal challenge, constitute disclosure. The audience never processes it. The defense always exists.

By the time the three seconds of disclosure pass, the viewer is watching a convincing synthetic human make eye contact with the camera and speak with the cadence of a candidate addressing voters. The brain is already applying the same trust heuristics it applies to video footage of real people: Is this person nervous? Do they seem to believe what they’re saying? Does their face match their words?

Disclosing that content is AI-generated before the viewer’s cognitive engagement with it begins would be disclosure. Displaying it briefly in the same place legal boilerplate lives is something else.

The Pattern

The Talarico ad is not an isolated case. The 2026 midterm cycle has produced a documented spread of AI-generated attack content across races at the federal, state, and local level, with different actors, different tactics, and different cover stories.

In Kentucky’s 4th congressional district, a super PAC ad depicted Republican congressman Thomas Massie in a fabricated “throuple” with Representatives Alexandria Ocasio-Cortez and Ilhan Omar. The ad included AI-generated video of the three together in domestic scenarios — dining, holding hands, checking into a hotel — that never happened. The cover story: satire. The framing: political commentary, not a factual claim.

In Maine’s U.S. Senate race, the National Republican Senatorial Committee ran an ad showing an AI-generated image and voice of a shirtless Graham Platner reciting Reddit posts he had written ten to fifteen years earlier. The posts were real; the synthetic figure delivering them was not.

In New Jersey’s 7th congressional district, a shadowy PAC called Real Change — which had reported no donors and no prior expenditures before suddenly spending hundreds of thousands of dollars on anti-Democrat ads — distributed mailers showing AI-generated images of candidate Rebecca Bennett standing among ICE officers in riot gear. The mailer included no disclosure of any kind. Bennett won the primary anyway. The PAC’s Republican fingerprints were noted by observers but never confirmed publicly.

What connects these cases is not a single party, a single tactic, or a single level of government. It is the same structural question repeated across different races with different answers: Do I have to tell anyone this is fake? And if I don’t, what happens?

The answer, in most of these cases, was: no, and nothing.

Two Layers

The NRSC approach to the Talarico ad is worth isolating because it is more sophisticated than the Kentucky “throuple” fabrication, and the sophistication is itself a manipulation choice.

The Kentucky ad manufactured scenarios that never happened. That’s the obvious kind of fake — entirely invented, easily verified as false, defensible only as “satire.” It requires a viewer to assess whether a fictional scenario is being presented as real.

The NRSC ad did something harder to dispute. It took Talarico’s actual past statements — real tweets, real positions, real words from real years — and delivered them through a synthetic version of the candidate. The content is substantially true. The vessel is entirely false.

This distinction matters because of how people evaluate attacks on political candidates. A fabricated scenario can be checked against reality and debunked. Real statements delivered by a synthetic persona occupy a different epistemic position: the statements are real, so verification confirms them. The delivery is fake, but the typical viewer is not equipped to separate the impact of the content from the impact of watching someone say it directly to a camera.

“In His Own Words” is the title of an ad in which none of the words came from Talarico’s mouth. The phrase is designed to signal authenticity — direct quotation, no editorializing, the candidate condemns himself. The synthetic face is designed to deliver that signal with the visual authority of actual video. The micro-disclosure is designed to make the legal file complete.

The manipulation operates across three registers simultaneously: the content (real), the delivery (fake), and the framing (authentic).

The Regulatory Vacuum

Twenty-eight states have enacted some form of law governing AI in political advertising. Most are untested in court. They are inconsistently written: some cover video only, some cover audio, some require prominent disclosure, some prohibit use in narrow circumstances without covering the broader category. No federal law exists. The Federal Election Commission has not acted. Democrats in the House have announced plans to pursue legislation if they retake power in 2027 — a contingency that does nothing for voters casting ballots in 2026.

The effect of this landscape is that campaigns and PACs can calculate their exposure precisely. In states with disclosure requirements, technically compliant disclosure — the kind that satisfies a legal check without informing a viewer — becomes the operational standard. In states without, no calculation is required.

The NRSC’s tiny-text approach reflects a reasonable legal read of a world where the rules are thin and contested. No court has ruled that a three-second small-print disclosure is insufficient. The standard has not been set. Until it is, the standard is self-set, and the incentive is to set it as low as possible while maintaining the ability to say you met it.

Influence Tactics Breakdown

Synthetic Authority Construction. Using a convincing AI-generated likeness of a real person to deliver political content creates the impression of first-person confession. Watching a synthetic Talarico make eye contact with the camera and speak triggers the same cognitive processing as watching the real Talarico do so — processing that a text quotation of the same content does not trigger. The manipulation is in the medium, not just the message.

Compliance Laundering. A disclosure that technically satisfies no existing legal requirement — because the requirement doesn’t exist — creates a paper trail without creating transparency. If challenged, the NRSC can point to the disclosure and demonstrate compliance with the standard they invented. The disclosure is addressed to future legal proceedings, not present viewers.

Authentic Content, Synthetic Delivery. Anchoring fabricated presentation in real source material (actual tweets, actual positions) makes the manipulation harder to contest. “Those are his real tweets” is true. “That is not him saying them” is also true. These two facts coexist in the same ad, and the format is designed to obscure the relevance of the second.

Regulatory Vacuum Exploitation. The proliferation of AI attack ads across 2026 midterm races follows a consistent pattern: actors test what can be done in the absence of enforceable standards, establish operational norms, and normalize practices before regulation catches up. By the time legislation passes, the behavior is established, the templates exist, and the question shifts from “should this be allowed?” to “how do we unwind what already exists?”

Satire Cover. The Kentucky ad demonstrates the alternate version of the same exploit: label fabricated scenarios as satire, and the content shifts from falsifiable claims to protected political commentary. “We didn’t say this happened — we imagined it” is not a defense a responsible actor offers, but it is a legally useful one in the absence of clear rules about what fabricated electoral content is permitted to look like.

What Viewers Actually Experience

There is a useful test: describe the Talarico ad to someone who hasn’t seen it. Tell them a national political organization created a synthetic version of a Senate candidate and had that synthetic version deliver the candidate’s real past statements plus some invented ones. Tell them the disclosure lasted three seconds in small text.

Then ask them what they think a viewer who watches the ad experiences.

Most people can answer this accurately. The description of the manipulation is not mysterious. What is missing — what the regulatory vacuum creates — is any mechanism by which the thing that’s clearly problematic when described is prohibited in practice.

The backfire effect in the Talarico case is documented: the ad generated significant attention, coverage of its nature as a deepfake, and, reportedly, increased sympathy for the candidate. The NRSC ad failed on its own terms. But “failed” here means “did not work in this specific instance with this specific candidate who had sufficient visibility for the coverage to spread.” The structure that produced the ad — national party organization, no disclosure requirement, minimal compliance theater — remains intact. A candidate with less visibility in a less-watched race faces the same ad with fewer defenses.

The question for the 2026 midterms is not whether deepfake attack ads work. Some do. Some don’t. The question is whether the people producing them face any cost for producing them, regardless of outcome.

Currently: they don’t.

The Compliance Question

There is a distinction between two versions of an AI-generated political ad that looks identical from the outside.

Version one: an actor creates an AI-generated attack and knows no disclosure is required, includes none, and relies on the content being accepted at face value.

Version two: an actor creates an AI-generated attack, includes a minimal disclosure that technically satisfies the appearance of compliance, and relies on viewers not processing the disclosure in time to affect their reception of the content.

Version two is the more sophisticated manipulation because it has pre-answered the accountability question. If called out, it points to the disclosure. The content is misleading. The process was transparent. These two things coexist.

The regulation debate focuses on version one — on establishing disclosure requirements. It is worth noting that version two already exists, is already being practiced, and already demonstrates that disclosure requirements can be met in a way that provides no protection against the manipulation they’re designed to prevent.

A disclosure requirement that isn’t designed around how human cognitive processing actually works is not a disclosure requirement. It is a compliance box. And compliance boxes, in political advertising, will always be checked at the minimum cost necessary to check them.

This article is part of Decipon’s Manipulation Breakdowns series, examining specific influence operations through the Influence Tactics Protocol.

Sources: