A Republican congressional candidate in Georgia published an AI-generated deepfake of Democratic Senator Jon Ossoff as part of a Senate primary campaign. A Massachusetts gubernatorial candidate ran an undisclosed AI voice clone of the sitting governor, attacking her own record. And a Democratic mayoral candidate in New York City briefly posted, then deleted, a deepfake ad targeting his opponent with racial stereotypes.

These aren’t isolated provocations from fringe actors. They’re opening moves. The most technologically disruptive midterm cycle in American history—and the regulatory infrastructure meant to contain it is, at best, a patchwork quilt with very large holes.

4×
More persuasive than traditional political ads — AI chatbots in controlled studies across three countries
Lin et al., Nature, Dec. 2025
28
U.S. states with deepfake disclosure laws as of February 2026, per NCSL data
STACK Cybersecurity / NCSL
26.1
Point shift toward agreement in optimized AI persuasion model — initially disagreeing participants, 77,000-person UK study
Science companion study, Dec. 2025
~100
AI-related measures passed across 38 states in 2025, per NBC News reporting on NCSL data
NBC News, Jan. 2026

Why 2026 Hit a Crisis Point — Not 2024, Not 2028

Three things converged. First, the cost of generating convincing synthetic media collapsed. The deepfake of President Biden used in the 2024 New Hampshire primary robocall — which federal investigators estimated reached between 5,000 and 25,000 voters — was made not by a sophisticated adversary but, as Senate testimony revealed, by a street magician. A street magician. Tools that required specialized technical skills in 2022 are freely available online now.

Second, the Federal Election Commission remains in partisan deadlock, unable to issue guidance on AI in political advertising. Not slow. Not cautious. Blocked.

Third, Congress passed the “One Big Beautiful” bill in May 2025, containing a provision that would impose a ten-year moratorium on state-level AI laws. The states that had acted most aggressively — California, Minnesota — are now facing both federal preemption risk and active First Amendment litigation. That’s the regulatory ground in early 2026: ambitious at the state level, paralyzed federally, constitutionally contested throughout.

“The question for 2026 isn’t whether AI will influence elections. It’s whether any institution can move fast enough to matter.”

Editorial synthesis — sources: Lin et al. (Nature, 2025), Brennan Center (2025), Cornell JLPP (2025)

As Campaign Now documented, a candidate targeted by a deepfake must spend finite campaign resources not on policy debate but on debunking fabrications. By the time a correction circulates, the fabrication has already shaped perception. The 2026 cycle didn’t create this problem. It’s the first cycle where the problem operates at scale.

The Persuasion Research: Larger Effects Than Anyone Wanted to Find

Here’s the thing about the empirical picture: it’s more complicated than either the alarm or the reassurance camps want to admit. Three peer-reviewed research programs, each with different methodologies, produce findings that partially contradict each other. And all three matter.

The most consequential comes from a December 2025 study published in Nature by researchers at MIT, Cornell, and Carnegie Mellon. Peer-reviewed. DOI: 10.1038/s41586-025-09771-9 Lin et al. recruited 2,300 participants across three countries — the 2024 U.S. presidential election, the 2025 Canadian federal election, and the 2025 Polish presidential election — to converse with AI chatbots advocating for opposing candidates. The effect sizes: Trump supporters moved 3.9 points toward Harris on a 100-point scale after a single AI conversation. That’s roughly four times the measured persuasion effect of political advertisements in the 2016 and 2020 elections. A companion study published simultaneously in Science, involving nearly 77,000 participants in the UK across 700-plus political issues, found the most optimized persuasion model shifted initially disagreeing participants 26.1 points toward agreement.

These are not marginal effects.

Second-order mechanism — why this doesn’t self-correct

The chatbots in these studies persuaded primarily through volume of fact-adjacent claims, not psychological manipulation. But optimizing models for persuasiveness came at the direct cost of accuracy — the most persuasive models also spread the most misinformation. MIT Technology Review’s analysis noted that chatbots advocating for right-leaning candidates made more inaccurate claims than those advocating for left-leaning candidates in all three countries studied.

That finding doesn’t appear in the usual coverage. The persuasiveness-accuracy tradeoff is the mechanism by which false claims scale, not just a peripheral data point.

Two PNAS studies complicate the picture from the other direction. Hackenburg and Margetts (2024) deployed GPT-4 in real time to generate personalized political messages and found that microtargeted messages were not statistically more persuasive than non-targeted messages in aggregate — casting doubt on whether personality-based customization adds meaningful persuasion lift beyond the baseline persuasiveness of the LLM itself. A second PNAS study from May 2025 found similar results on microtargeting and interactive elaboration.

So: three components, and they don’t add up cleanly. Baseline LLM persuasiveness is large and documented. The personality-targeting persuasion premium is mixed and contested. Chatbot engagement rate in real political campaigns remains an open empirical question. Treating any one of them in isolation produces either excessive alarm or false reassurance. Most coverage picks one.

Study Method & Scope Key Finding Evidence Level ⚠ What This Doesn’t Establish
Lin et al., Nature (Dec. 2025) N=2,300 across 3 countries; randomized chatbot conversations; peer-reviewed AI chatbots ~4× more persuasive than traditional political ads; most persuasive models also spread most misinformation Strong Controlled lab conditions; real-world engagement rates with AI chatbots in live campaigns are unknown. Effect size may not hold at scale or with habituation.
Hackenburg & Margetts, PNAS (Oct. 2024) Live GPT-4 deployment; real-time personalized messaging; peer-reviewed Microtargeted messages not statistically more persuasive than non-targeted in aggregate Moderate Measures the targeting premium only — not baseline LLM persuasiveness. Cannot be used to argue AI chatbots are ineffective persuaders overall.
Kim et al., PNAS (Jan. 2026) Verified voting records + ad exposure + demographics; 2016 election data; peer-reviewed Targeted digital voter suppression ads disproportionately reached non-White voters in battleground states and measurably reduced turnout Strong for 2016; directional for 2026 2016 data. Technological capabilities have expanded significantly. Directional for 2026 only — requires updated replication to confirm current-cycle effects.
Simchon et al., PNAS Nexus (Feb. 2024) Randomized controlled trial; personality-based microtargeting with GPT-4; peer-reviewed Microtargeting plus interactive elaboration showed no clearly superior persuasive effect vs. non-targeted Moderate Sample not drawn from contested-district populations. Results may not generalize to voters in high-stakes, high-information races.
Sources: Nature (2025), PNAS (2024, 2026), PNAS Nexus (2024). Studies measure different dependent variables; direct comparison requires caution. Evidence levels: Strong = consistent findings from robust peer-reviewed studies with large samples or established data linkage. Moderate = solid findings with significant generalizability limits or replication gaps. Directional = suggestive but insufficient for confident projection to current context.

What Campaigns Are Actually Using AI For

If you’re running a down-ballot campaign right now, you’re almost certainly being pitched AI on three grounds: it cuts the cost of voter outreach, it personalizes fundraising appeals, and it helps smaller campaigns punch above their weight against incumbents. That pitch is largely accurate. It papers over a set of ethical questions most campaigns have not formally resolved.

The American Prospect’s October 2025 analysis documented three distinct modes. Professional campaigners — ad buyers, fundraisers, digital strategists — treat AI as an efficiency tool: automating personalized email and text drafting, optimizing ad targeting, replacing some functions previously handled by paid consultants. Community organizers are deploying it to identify and mobilize latent supporter networks. And individual citizens are using generative AI to both create and amplify political content — a category that includes the deepfake attack ads circulating in competitive races.

The production problem campaigns are learning the hard way: the line between the second and third categories is porous. A Pennsylvania Democratic congressional candidate’s AI robocaller, “Ashley,” disclosed its AI nature at the top of each call — a transparency measure the campaign viewed as a feature. Same technological infrastructure, deployed without disclosure or with adversarial intent, is indistinguishable from the voter suppression operations that Kim et al.’s January 2026 PNAS study documented targeting non-White voters in battleground states in 2016. First study to use verified voting records — not self-reported data — to demonstrate that targeted digital voter suppression ads measurably reduced turnout. Tier 1 evidence.

Cross-source synthesis — not present in any single cited source

The infrastructure that makes ethical campaigns more efficient is identical to the infrastructure that makes voter suppression operations harder to trace. The Lin et al. persuasion effects, the Kim et al. turnout suppression mechanism, and the platform self-regulation failure documented by the Georgetown Center for Cyber Politics together imply a specific 2026 risk that none of the three sources names directly: as AI persuasion tools become cheaper and more accessible, the asymmetry between use for outreach and use for suppression disappears — and the populations most exposed to suppression in the documented historical cases are the same ones most likely to be targeted in competitive 2026 Senate and gubernatorial races.

“The AI infrastructure that makes ethical campaigns more efficient is identical to the infrastructure that makes voter suppression operations harder to trace.”

Editorial synthesis — sources: Kim et al., PNAS (2026); Lin et al., Nature (2025); American Prospect (2025)

How Effective Are the Existing Rules? (Short Answer: It Depends What You Mean by “Rules”)

As of January 2026, 28 states have enacted laws specifically addressing deepfakes in political communications, per STACK Cybersecurity’s tracker citing NCSL data. Most require disclosure rather than prohibition. Montana and South Dakota enacted deepfake disclosure requirements in 2025 that take effect for the 2026 midterms. NBC News reported in January 2026 that 38 states passed roughly 100 AI-related measures in 2025 — a volume that signals legislative urgency, if not coherence.

California’s approach — the most aggressive — illustrates the ceiling on state action. AB 2839, enacted September 2024, was partially struck down by a federal judge in August 2025, with key provisions found to conflict with Section 230 of the Communications Decency Act and potentially unconstitutional under the First Amendment. Cornell Law’s Journal of Law and Public Policy 2025 analysis of the litigation landscape concluded that courts remain skeptical of sweeping prohibitions on political deepfakes precisely because they risk chilling satire and parody — categories with long First Amendment protection. Minnesota’s law has faced a challenge from X (formerly Twitter) on Section 230 grounds, with early rulings unfavorable to the state.

The FCC acted within its narrower lane: in February 2024, the agency ruled that AI-generated and voice-cloned audio qualifies as an “artificial or prerecorded voice” under the Telephone Consumer Protection Act, requiring prior express consent for robocalls and robotexts. That ruling covers telecommunications. It doesn’t cover AI-generated video, manipulated images, or synthetic content distributed on social media — which is where most of the contested 2026 material is actually circulating.

The Structural Problem: Why Disclosure Rules May Not Be Enough

Disclosure requirements operate on a theory of voter cognition: if people know content is AI-generated, they’ll discount it appropriately. That theory may underestimate how persuasion actually works.

The Lin et al. Nature study is instructive here — all participants in those experiments were explicitly told they were talking to an AI. The persuasion effects remained substantial. The most persuasive AI chatbots didn’t hide their nature. They just outargued the human’s prior beliefs with a higher volume of fact-adjacent claims than a human conversation partner would typically deploy. Disclosure didn’t neutralize the effect.

GWU’s Peter Loge, director of the Project on Ethics in Political Communication, has articulated this clearly: campaign AI ethics are not categorically different from campaign ethics in general. The injunction to persuade rather than deceive predates AI by millennia. What AI changes is scale, cost, and the rate at which a single bad actor can deploy deceptive techniques across an electorate. A street magician can produce a presidential deepfake robocall. That was inconceivable five years ago.

The deniability problem — second-order mechanism

The Brennan Center for Justice’s analysis flags a particular risk that disclosure requirements don’t address: heightened public awareness of AI in campaigns gives politicians an incentive to deny the authenticity of real content — to claim genuine video or audio is a deepfake when it isn’t.

If voters can’t reliably distinguish synthetic from authentic, the mere existence of deepfake technology becomes a plausible deniability tool for candidates confronted with legitimate evidence of their own conduct. This is not a hypothetical. It’s a documented structural consequence of widespread synthetic media.

Where This Goes: Three Scenarios for the 2026 Cycle and Beyond

Trajectory 01 — Most Likely Near-Term

Platform Self-Regulation as De Facto Governance

In the absence of federal standards, Meta, YouTube, and X become the primary arbiters of what AI-generated political content is permitted and removed. That arrangement concentrates enormous power in private companies with their own political and commercial incentives. Already producing inconsistent outcomes: the same deepfake content persists on one platform while being removed from another. The Georgetown Center for Cyber Politics, as reported by Route Fifty, has flagged that fact-checkers and watchdog organizations are structurally overmatched by algorithmically distributed content, particularly through encrypted apps outside the major moderation frameworks.

Trajectory 02 — Underreported Risk

AI Persuasion Becomes a Structural Partisan Asymmetry

Time’s October 2025 analysis observed that conservative and right-aligned campaigns adopted AI-generated political imagery most aggressively in global 2024 elections. If AI-powered voter outreach and persuasion tools are adopted more heavily by one party’s campaigns — due to differential tech fluency, resource distribution, or ideological affinity — the persuasion effects documented in Lin et al. could produce systematic electoral advantages compounding across competitive districts. The most impactful uses of AI in the 2026 cycle, The American Prospect warned, may not be visible until 2027 or later.

Trajectory 03 — The Most Underreported

Participatory AI: Using the Technology the Other Direction

D.C. Mayor Muriel Bowser’s partnership with MIT and Stanford to use AI-based public deliberation tools in city policymaking offers a proof of concept: AI can be used to identify common ground across a large population, not just to target individual vulnerabilities. Tools like Decidim and Pol.Is have been deployed in European policy contexts. Whether any 2026 U.S. candidate deploys them at meaningful scale remains to be seen — but the capability exists and the model is documented.

What You Actually Do With This

For campaigns operating in competitive Senate and gubernatorial races over the next nine months: the most important practical question is not whether to use AI but what internal governance structure you have for approving AI-generated content before it’s published.

The Massachusetts governor’s race has already demonstrated that an undisclosed AI voice clone generates backlash and legal exposure regardless of its technical sophistication. The Georgia Senate race has demonstrated that a labeled or unlabeled deepfake attack ad can generate both media attention and opposition fundraising material simultaneously. These are not symmetric risks — they cut in multiple directions at once, and campaigns without a content approval workflow specific to AI-generated material are navigating them without a map.

The voters most exposed to AI-driven voter suppression in documented historical cases were non-White residents of battleground state counties. People for whom the gap between “uncertain about whether to vote” and “decided not to vote” can be manufactured by a well-targeted ad delivered at the right moment. The infrastructure to do that in 2026 is cheaper, more capable, and less regulated than it was in 2016. Courts are striking down the laws meant to constrain it. The FEC can’t agree on a framework.

We’re not watching a problem approach. We’re watching a governance system fail in real time, one competitive district at a time.

For Campaigns: The Practical Decisions Right Now

For: Campaign Managers & Senior Strategists

Build the Approval Gate Before You Need It

The reframe isn’t “should we use AI” — it’s “who owns the liability decision when an AI-generated asset goes wrong.” Right now that answer in most campaigns is nobody, which is why the Massachusetts governor’s race and the Georgia Senate race produced the outcomes they did.

What you do: Designate a named individual — not a department, a person — responsible for approving AI-generated content before publication. That approval process must include: (1) disclosure status documented, (2) source attribution for any AI-assisted research claims, (3) sign-off that the content doesn’t misrepresent the opposing candidate’s record using fabricated audio or visual. It doesn’t need to be elaborate. It needs to exist and be enforced.

Here’s what’s going to stop you: the production speed advantage of AI tools creates pressure to skip approval steps. The campaigns most likely to face legal exposure are the ones that moved fastest without a process. The Massachusetts case is the canonical example — the speed was not worth the outcome.

Stop doing this: Don’t treat disclosure as a legal checkbox you add after the asset is created. Disclosure status has to be part of the production brief, not the review. A post-hoc disclosure process fails faster than it gets implemented.

For: Civic Technologists & Policy Researchers

The Research Gap That Actually Matters for 2026

The missing empirical link in everything here: real-world chatbot engagement rates in live political campaigns. The Lin et al. Nature study establishes large persuasion effects under controlled conditions. The Hackenburg & Margetts PNAS study establishes that microtargeting adds limited premium over baseline LLM persuasiveness. What neither establishes is what percentage of voters in competitive 2026 races will actually interact with an AI persuasion agent — versus seeing a static deepfake ad or receiving a targeted text.

What you do: The Kim et al. 2026 PNAS study methodology — linking verified voting records to documented ad exposure — is the model. Applying that methodology to AI chatbot and robocall exposure in 2026 competitive races would close the most important empirical gap in the current literature. The data to do this won’t exist until after November 2026, but the research design can be built now.

Here’s what’s going to stop you: the advertising exposure data is often held by the platforms. Meta and Google both have mechanisms for academic data access, but the timelines are slow relative to the election cycle. Planning that access now — not after the election — is the practical constraint.

Stop doing this: Stop treating the Lin et al. persuasion effects as directly translatable to real-world electoral impact without the engagement-rate bridge. The effect sizes are large under controlled conditions. That doesn’t resolve what happens when voters encounter AI persuasion in the wild, where they have competing information sources, varying motivation, and — increasingly — skepticism toward political content of any kind.