AI in Elections and Democratic Institutions

"AI in elections" refers to the use of generative or algorithmic AI systems to influence electoral outcomes, voter behavior, or democratic discourse. By 2026, generative AI is a standing feature of political communication in major democracies, and the policy response is fragmented across multiple regulators with no comprehensive US federal statute. Lines of authority are split among the Federal Election Commission (paid political advertising), the Federal Communications Commission (broadcast and robocalls), state election laws (deepfake-in-campaign statutes), private platform rules (Meta, Google, TikTok), and the EU Digital Services Act.

Scope

The category covers several distinct uses of AI in the electoral context:

• Synthetic political media — AI-generated audio, video, or images depicting candidates or events.
• Generative content at scale — AI-produced social-media posts, emails, ads, or comments.
• AI-enabled microtargeting — algorithmic audience segmentation and message customization.
• Robocall and voice-clone scams — AI-synthesized voices impersonating candidates or officials.
• Election integrity systems — AI tools deployed by platforms or governments to detect disinformation, identify coordinated inauthentic behavior, or manage content moderation.
• AI-driven voter interaction — chatbots answering election-related queries.

Structural dimensions

Three distinctions in existing law and platform practice shape how AI in elections is governed.

The first is paid versus organic speech. US election law distinguishes paid political advertising (the FEC regulates disclosure; platforms apply ad policies) from organic content (largely unregulated). AI narrows this distinction, because organic content generated cheaply at scale can carry the strategic weight of a paid campaign, and existing frameworks largely do not reach organic AI content.

The second is platform versus state. Private platforms (Meta, Google/YouTube, TikTok, X) maintain their own political-content and AI-content rules, often more restrictive than any applicable law. In the US, the effective regulator of AI political content is therefore private, an arrangement that allows rules to change faster than statutes but also allows them to be withdrawn, as Meta's 2024 retreat from election-integrity infrastructure illustrated.

The third is domestic versus foreign-origin. Detection and attribution of AI-generated political content is difficult. The Russia/China/Iran disinformation frame still applies, but AI tooling erodes the cues, such as translation errors and image inconsistencies, that human reviewers historically relied on.

Documented incidents

Biden New Hampshire robocall (January 2024)

An AI-synthesized voice impersonating President Biden made robocalls to New Hampshire voters before the state's 2024 primary, instructing Democrats to "save your vote" for November rather than vote in the primary. The consultant responsible, Steve Kramer, working with a Democratic challenger campaign, was later indicted. The FCC imposed fines and, in a February 2024 ruling, declared AI-voice robocalls illegal under the Telephone Consumer Protection Act (TCPA). The incident is the most-cited case of voice-clone disinformation in US elections and directly drove FCC rulemaking. Voice cloning is the primary technical mechanism behind documented election-interference incidents, including this robocall and the Slovak Šimečka audio (AI Voice Cloning).

California local election influence attempt (June 2026)

On July 13, 2026, California local election officials said AI was used in an attempt to influence a June 2026 election, as officials and researchers described AI as giving election conspiracy theories a new playbook heading into the 2026 midterms (Source: politico.com).

Systematic testing of chatbots and generators

The first systematic published test of how consumer AI systems handle election disinformation is the Brennan Center for Justice expert brief "Does AI Fight or Fuel Election Disinformation?", published August 11, 2026 by Abdiaziz Ahmed, Alex Brunet, Mark Krass, Lawrence Norden, Marcelo Agudo, Owen Doyle and David Evan Harris, based on tests run between February and August 2026 (Source: brennancenter.org; techpolicy.press).

On the question of whether chatbots repeat conspiracy theories, the finding is uniformly negative: every one of the six systems tested — ChatGPT, Gemini, Grok, Claude, Perplexity and DeepSeek — disputed every election conspiracy theory put to it, including under repeated questioning from users posing as sympathetic to those theories. The accuracy findings are weaker. Half of all responses contained an inaccuracy or a bad citation, one in three contained a factual error, and one in three contained broken links or misleading citations.

On generation rather than answering, the results diverge by system. Asked to draft a full set of image-generation prompts for fabricated election claims, most chatbots refused; Grok produced the complete set, with its chain of thought recording that "election misinformation not listed as disallowed activity." Safeguards also applied inconsistently across tiers of the same product: Meta AI's versions returned contradicting acceptances and denials to the same prompt 67 percent of the time, ChatGPT 28 percent, Grok 22 percent and Gemini 10 percent, while Runway Gen-4 and Flux.2 never differed between versions because neither version rejected a single prompt.

On detection, asked to classify 14 AI-generated images, Gemini and Meta AI correctly identified most as synthetic while ChatGPT, Grok, Claude and Perplexity each identified four or fewer. Reruns after California's AI Transparency Act took effect on August 2, 2026 showed little improvement except for Gemini, which had switched to reading embedded provenance data — a result that ties detection performance to the provenance-metadata track rather than to image forensics.

AI as a campaign issue

Separate from AI's use in political communication is its emergence as a subject of it. Shira Ovide, Clara Ence Morse and Kevin Schaul published an analysis of candidate campaign websites on August 14, 2026 finding that more candidates in the 2026 US midterm races mention artificial intelligence or data centers than mention Israel or manufacturing (Source: washingtonpost.com). Only the article's opening was retrievable and the underlying counts were not, so the comparison is recorded as the authors state it without the supporting figures. The finding pairs AI with data centers as a single campaign topic, which is consistent with the local-siting disputes tracked on Data Center Siting / AI Power Politics rather than with the model-safety debate.

The same pairing appears from the campaign-strategy side. A National Republican Senatorial Committee memorandum of August 18, 2026 describes data-center opposition, rather than any AI-generated content, as the most damaging line of attack in the Ohio Senate race, reports three television advertisements and more than 6,000 points of broadcast messaging on the subject from Sherrod Brown, and calls it "a sleeper issue for the entire election cycle." The memorandum's request is directed not at platforms or regulators but at the companies building the projects, which it asks to answer "who benefits, who pays, and why a community should want one" (NRSC — Ohio Data Center Risk (memorandum, August 18, 2026)). The electoral question it raises is about AI infrastructure and its local costs, not about synthetic media — a distinct axis from the rest of this page. See Data Center Siting / AI Power Politics.

Advice to candidates followed. A memorandum dated September 14, 2026 from Republican pollster Tony Fabrizio to "GOP Candidates & Policymakers" states that "unqualified support for building more AI data centers is a losing position," reporting survey results in which voters opposed more construction 65% to 24% while 61% preferred construction with conditions on power costs, ratepayer protection and water. It advises candidates to pair any statement of support with enforceable conditions and to avoid the phrase "build data centers," which it says "can be clipped and put in an attack ad." The memorandum states that its survey did not test effects on vote share, and it does not name the survey's sponsor, field dates or sample (Fabrizio Lee — Neutralizing Political Attacks on AI Data-Center Development Through Enforceable Conditions (memorandum, September 14, 2026)).

Surveillance technology has opened a third campaign line on the same axis. Abdul El-Sayed, the Democratic nominee for Senate in Michigan, posted a video on August 19, 2026 tying his Republican opponent, former Rep. Mike Rogers, to Flock's license plate reader cameras, and Sen. Bernie Sanders wrote the day before that "Flock has installed nearly 100,000 cameras across America to track your every movement." Rep. Tim Burchett and two other House Republicans introduced H.R. 9800 in the 119th Congress to bar federal purchases of Flock cameras or similar technology, making the issue bipartisan in its opposition if not in its framing (Source: axios.com). See AI and Surveillance, AI Backlash.

Candidates have begun making formal commitments on the same axis. More than 15 politicians running for office across the United States have signed the AI Pact, a set of five promises covering data centers and AI, in an account published August 26, 2026. The promises include advocating mandatory safety reviews for AI models, supporting policies that would create an AI dividend for workers, ensuring consumers and companies can sue over harms caused by models, and supporting policies banning tax breaks, backroom deals and excessive energy usage for data centers; it does not call for a construction moratorium. Cofounder Daniel Lobo-Lewis said the pact is an offshoot of the Political Integrity Project, is funded by a progressive political action committee, and receives no direct funding from AI safety groups or technology companies. Every current signatory is a Democrat except Dan Osborn, the independent Nebraska Senate candidate; others include Representative Justin Pearson of Tennessee and Texas gubernatorial candidate Gina Hinojosa (Source: wired.com). The pledge text itself has not been retrieved. It sits opposite the Leading the Future super PAC network on the same 2026 cycle. See Data Center Siting / AI Power Politics.

Regulatory response

FEC regulation of AI in political ads

The Federal Election Commission launched rulemaking on "AI deepfakes in political advertising" in 2023–2024 but declined to issue a rule in August 2024, on the grounds that existing anti-fraud provisions were sufficient and that it lacked authority to regulate AI content specifically. The result is that there is no federal paid-political-ad-specific AI rule. The FEC's position has been criticized as an abdication; defenders argue the Commission lacks statutory authority and First Amendment clearance.

Earlier legal analysis had anticipated a narrow role. Robert Chesney and Danielle Citron (2019) noted that the FEC does not regulate the truth of campaign statements and was unlikely to assert such jurisdiction given First Amendment obstacles, that major online advertising platforms were then outside its disclosure rules, and that those rules are candidate-specific and do not reach issue ads. They saw room for a requirement that creators of election deep fakes disclose that the content is fake, since the Supreme Court has upheld campaign-speech rules requiring truthful disclosure of a communication's source (Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security (Chesney & Citron, California Law Review, 2019)).

FCC robocall rules

The FCC, which regulates telecommunications under the TCPA, acted after the Biden robocall. A February 2024 Declaratory Ruling held that AI-generated voices in robocalls constitute "artificial voice" under the TCPA, subjecting them to the statute's prior-express-consent requirement and opening the door to civil penalties. The FCC has subsequently proposed rules on AI disclosure in political robocalls and texts.

Election-security coordination (2026 cycle)

On July 13, 2026, Rep. Mike Lawler (R-NY) urged CISA to coordinate with the Federal Election Commission and the Justice Department on mitigating AI-enabled threats to the 2026 election cycle (Source: insideaipolicy.com).

Sen. Adam Schiff and Rep. Ro Khanna reintroduced the AI Ads Act on Monday, July 27, 2026 — a measure addressing the gap the FEC declined to fill in 2024. The available reporting names the sponsors and the reintroduction date but does not set out the bill's provisions or its number (Source: nextgov.com).

State deepfake political ad laws

By 2026, a majority of US states have enacted deepfake political ad statutes, typically requiring disclosure that synthetic media is AI-generated, often with a window (for example, 60 or 90 days before an election) during which undisclosed synthetic depictions of candidates are actionable. Texas, California, Washington, Minnesota, and Michigan have among the most developed regimes. The statutes vary on whether they require disclosure or prohibit use outright; whether they apply only to candidate campaigns or also to PACs and independent expenditures; whether they provide a private right of action (some do, most do not); and First Amendment vulnerability, with most statutes narrowly tailored to avoid constitutional challenge but a handful enjoined.

Platform election integrity policies

Major platforms adopted election-AI rules in 2023–2024, followed by partial retreat:

• OpenAI published election integrity principles in January 2024, barring use of its tools to build applications that discourage voting or misrepresent the voting process, and updated ChatGPT to direct election queries to CanIVote.org in the US.
• Anthropic published an election-integrity blog post in 2024, committing to Acceptable Use Policy enforcement against voter suppression, candidate impersonation, and disinformation production, and piloted the Prompt Shield election-query safeguard.
• Google restricted Gemini from answering most election-related queries in 2024, a move criticized as over-cautious but which reduced the error surface.
• Meta retrenched in 2024–2025, dissolving much of its election-integrity team and relaxing political-content rules; civil society has cataloged the retreat as a major regression.
• X (Twitter) under Musk rolled back most of the pre-2022 election integrity infrastructure.

EU Digital Services Act

The EU AI Act interacts here with the Digital Services Act (DSA), which predates it. The DSA requires very large online platforms (VLOPs) and search engines (VLOSEs) to conduct systemic-risk assessments covering electoral processes and civic discourse, to publish risk-mitigation measures, and to permit independent audit. The European Commission used DSA powers in 2024 to open formal proceedings against TikTok (over its handling of Romanian elections) and X (over content moderation and advertising transparency). As of 2026, the DSA is the most active regulatory instrument in AI-elections governance globally, outpacing US federal action.

Microtargeting

AI-driven audience segmentation allows political campaigns to produce and deliver custom messaging at scale. Empirical research (Kalla and Broockman 2018) suggests microtargeting's direct persuasion effects are modest, but AI reduces the marginal cost of message production toward zero, shifting the equilibrium even if per-message effects are small. The EU's Political Advertising Regulation, applicable from October 2025, restricts AI-powered microtargeting in political ads, requiring explicit consent for sensitive-data-based targeting, which is the most restrictive such measure currently in force.

Policy responses at a glance

Debates and positions

Coverage of AI in elections has been split between accounts that emphasize direct manipulation through synthetic media and accounts that emphasize broader erosion of trust. Both strands appear in Chesney and Citron's 2019 article on deep fakes. On manipulation, they argued that elections present "decisional chokepoints," "narrow windows of time during which irrevocable decisions are made," so that a fake released with enough time to circulate but not to be debunked could affect an outcome; they used the 2017 Russian operation against Emmanuel Macron's campaign, which combined stolen and altered documents with a poorly timed release, to illustrate what a better-timed operation using deep-fake video or audio might achieve. They judged criminal law unlikely to deter the most capable actors, foreign intelligence services, and recommended a new national emergency under the International Emergency Economic Powers Act to allow sanctions against foreign attempts to inject false information into US political dialogue. On trust, the same article introduced the liar's dividend (Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security (Chesney & Citron, California Law Review, 2019)). Nathaniel Persily's essay "Misunderstanding AI's Democracy Problem" in The Digitalist Papers argues against deepfake-election panic, locating the more durable harm in trust erosion rather than direct persuasion, a framing connected to the liar's dividend concept introduced by Chesney and Citron. Taiwan's Information Integrity Alignment Assembly (March 2024) has been cited as a case study of a participatory approach to AI-elections governance (Taiwan Ministry of Digital Affairs (moda)).

Relationships

• instance-of: AI governance via sectoral regulation
• related: AI and Content Moderation — adjacent concept of AI-in-platform-governance
• related: AI and Authoritarianism — election-integrity concerns in authoritarian-leaning regimes
• related: AI and the First Amendment — constitutional constraints on deepfake statutes
• related: EU AI Act (Regulation 2024/1689) — AI Act intersects with DSA on elections
• related: Techno-Federalism: How Regulatory Fragmentation Shapes the U.S.-China AI Race — state-level fragmentation of deepfake regimes
• related: AI and Surveillance — microtargeting as surveillance infrastructure
• related: AI Compliance Industry / Regulatory Fragmentation — multi-regulator, multi-jurisdiction burden
• related: Liar's Dividend — primary election-democracy concept; trust erosion versus direct persuasion
• related: AI Content Provenance — embedded provenance data as the mechanism behind the one detection improvement measured after California SB 942 took effect
• related: Data Center Siting / AI Power Politics — AI infrastructure as a campaign subject, the second axis on which AI enters elections
• related: Fabrizio Lee — Neutralizing Political Attacks on AI Data-Center Development Through Enforceable Conditions (memorandum, September 14, 2026) — pollster's messaging memorandum to Republican candidates on data centers
• related: NRSC — Ohio Data Center Risk (memorandum, August 18, 2026) — campaign-committee primary text treating data centers as the decisive issue in a Senate race
• related: The Digitalist Papers (Stanford, Volumes 1–2) — Persily's "Misunderstanding AI's Democracy Problem" essay critiques deepfake-election panic
• related: Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security (Chesney & Citron, California Law Review, 2019) — decisional-chokepoint analysis of election deep fakes; FEC and IEEPA responses; origin of the liar's dividend
• related: Taiwan Ministry of Digital Affairs (moda) — Taiwan Information Integrity Alignment Assembly (March 2024) as case study of participatory approach to AI-elections governance
• related: AI Voice Cloning — voice cloning is the primary technical mechanism behind the documented election-interference incidents (NH Biden robocall, Slovak Šimečka audio)