AI in the 2026 Midterms: Deepfakes, Chatbots, and Election Security

Artificial intelligence has picked up several jobs in the 2026 midterms: campaign assistant, voter guide, administrative aide, and occasional impersonator. Its qualifications are still being checked, mostly because it has a habit of hallucinating facts like a sleep-deprived intern.

The US general election is Tuesday, Nov. 3, 2026, with early and mail voting already underway in many states. As Americans choose candidates for Congress and state and local offices, AI is helping shape the messages they encounter, the answers they receive, and some of the work behind their voting experience.

Campaigns are producing synthetic advertisements and trying to influence what chatbots say about candidates. Election officials are experimenting with AI for communications, staffing, and training. Technology companies are introducing safeguards, including an OpenAI election operations center announced less than a week before Election Day.

The evidence supports both practical benefits and serious concerns. It also demands some distinctions: a fabricated campaign video, an incomplete voter guide, an AI agent taking an unexpected action, and an attempted intrusion into a government website create different problems. Assessing preparedness means examining each one.

Campaigning gets cheaper, faster, and more synthetic

AI’s campaign work extends well beyond deepfakes.

It can help draft fundraising emails, summarize research, translate material, and produce graphics or video. The Washington Post reported that Democratic campaigns are using it behind the scenes even as many avoid conspicuously synthetic advertisements.

The advertising shift is substantial, proving that nothing gets politicians to adopt new tech faster than the promise of saving a buck.

In its Oct. 15 update, the Wesleyan Media Project identified 220 ads generated or enhanced with AI, associated with more than $137 million in advertising spending. That measures spending on the identified ads, rather than payments to AI companies or the cost of creating them. Researchers described the count as an underestimate.

Republican-affiliated sponsors accounted for 88% of that spending, according to Wesleyan. Only 27% of the identified ads contained an AI disclaimer. Those figures describe adoption and disclosure within the researchers’ sample; they do not establish that every AI-assisted advertisement was deceptive.

Some uses resemble conventional production work, such as creating generic background scenes. Others fabricate recognizable politicians’ words or actions. The New York Times documented synthetic attacks on Democratic candidates and Republicans competing in their own party’s primaries.

Campaign professionals disagree about where to draw the line. Conservative strategist Eric Wilson argues that obvious AI satire can function like political cartoons, lowering production costs and helping smaller campaigns compete. The central question is whether viewers understand that a scene is invented.

A cheaper production process offers no assurance that the message is true, just that it was spun up at a discount.

Campaigns are also campaigning to the chatbots

Campaigns increasingly have another audience: the AI systems voters consult.

Reuters reviewed a presentation proposing ways to influence chatbot answers about Mike Minogue, the Republican challenging Massachusetts Gov. Maura Healey. The pitch included hundreds of policy pages, placements on websites AI systems already cite, and content for platforms such as YouTube and Reddit.

The proposed service would monitor how chatbots answered political questions and adjust online material to improve the candidate’s presentation. It illustrates generative engine optimization, or GEO: making information more likely to appear in AI-generated answers. Reuters documented a sales pitch. However, it provided no proof that Minogue adopted it or that it successfully changed voters’ choices.

Similar services operate across political lines. CampSight, run by the progressive Run for Something Action Fund, monitors chatbot responses and advises campaigns on how to improve them. Publishing clear policy positions can help correct incomplete information. Paid placements and coordinated publishing also raise questions about whether an apparently independent answer is drawing on campaign-shaped material.

That concern extends beyond campaign websites. The New York Times and The Atlantic described a network of 37 purported local-news sites identified by security researchers that appeared designed for AI systems to ingest. Some pages favored Democratic candidates. The reporting did not establish that the candidates named on those pages operated the network.

The campaign trail now includes the source list beneath an AI answer. Readers need to know whether a citation leads to reporting, a candidate’s own position, paid advocacy, or a website whose ownership is unclear. AI summaries can even misdirect financial support, proving that even the machines get confused by campaign finance.

For example, a Google AI answer provided a misleading ActBlue link controlled by an unrelated PAC that initially diverted donations intended for Texas Senate candidate James Talarico to Senator John Hickenlooper’s reelection campaign, which then forwarded the misdirected funds back to Talarico, according to The Atlantic.

A confident voter guide can still leave things out

Deliberate chatbot use for election information remains a minority activity.

A Pew Research Center survey conducted Oct. 12–18 found that 8% of US adults said they were using chatbots for information about the 2026 elections. Among that group, 76% sought candidates’ issue positions, while 40% sought suggestions about their voting choices.

That does not capture every encounter with AI during an election. As The New York Times reported in its article, voters also encounter generated search summaries, synthetic campaign material, and platforms using automated moderation. Opening a chatbot is only one entrance.

The States United Democracy Center’s study of AI answers to voter questions illustrates a subtler reliability problem. Its primary testing collected 402 responses focused on Arizona, Michigan, and Pennsylvania in January and February. Researchers found no verifiable factual errors under their coding criteria, an improvement over preliminary testing in 2025.

Yet ChatGPT returned incomplete candidate lists in 88.9% of the gubernatorial queries. Responses also inconsistently directed users to official election websites. A guide can be correct about everyone it mentions and still leave candidates out, which is a fantastic strategy for avoiding lawsuits but a terrible one for educating voters. These findings describe particular products and testing conditions earlier this year, rather than every chatbot’s current performance.

Research also shows that design matters. Experiments published in Nature found that AI conversations advocating for candidates could shift participants’ preferences. A separate PNAS study found that a purpose-built voting advice chatbot improved young, politically unaligned participants’ knowledge while producing generally weak downstream effects on voting preferences. Those findings establish potential effects under experimental conditions, rather than a count of votes changed in the 2026 midterms.

Disinformation spreads faster than a correction can travel

As experts warned in The Guardian, “AI-driven disinformation [is] running rampant on social media before US midterms.”

Generation is only part of the problem.

Material must also reach an audience, and platforms’ disclosure, moderation, and recommendation systems influence that distribution. The Times’ reporting, drawing on research from the Institute for Strategic Dialogue, described uneven policies and limited visibility into how companies would handle emerging election risks. Companies including Google and Meta said they had extensive testing and election safeguards.

Corrections also struggle across languages and platforms. The Digital Democracy Institute of the Americas documented how a false rumor about Texas Senate candidate James Talarico picked up additional allegations as it spread through Spanish-language networks. That case illustrates how falsehoods mutate; the reporting does not establish that every version was AI-generated.

Trust can suffer in two directions. Fabricated recordings can persuade people that something happened, while awareness of deepfakes can lead them to dismiss authentic evidence as fake. Researchers call the latter the “liar’s dividend.” Reliable verification requires checking origins and corroborating evidence, rather than assuming that convincing footage is real or inconvenient footage is synthetic.

Election offices are finding useful, ordinary applications

The Election Assistance Commission’s March guide, AI in Action: Case Studies for Election Officials, describes officials using AI to draft a year’s social media calendar, deadline reminders, and public notices. Staff checked accuracy, adjusted terminology and tone, and added local context.

Other cases involved scheduling poll workers around skills, party-balance requirements, and preferences. An office also developed turnout forecasts using previous primaries, checking the method against known historical results before applying it to planning.

AI helped generate meeting summaries and training material, too. Errors required attention: a transcription inserted inappropriate words nobody had spoken, confused technical terminology, and mislabeled speakers. Staff verified motions and reviewed training against local procedures.

The guide recommends human review, approval before publication, documentation, and disclosure for public-facing AI-assisted content. It warns against entering individual voter registration information, security plans, privileged communications, or unpublished information into AI tools.

These examples show practical assistance rather than a national performance evaluation. The first draft may arrive faster. The deadline still has to be right, since “the chatbot ate my homework” doesn’t fly in election law.

OpenAI’s election war room has a specific assignment

OpenAI is establishing an Election Operations Center to identify potential misinformation and other attempts to disrupt elections through misuse of its tools, Politico reported on Oct. 28. The company said the operation would bring together hundreds of employees across teams, beginning this week and continuing until votes are counted.

Its work includes monitoring events and reviewing reports of unexpected model behavior or suspected policy violations. That gives OpenAI a mechanism for responding to problems involving its services. It does not give the company authority over election administration or visibility into every competing AI system.

OpenAI also says its Democracy Works integration supplies voting logistics across languages and its Associated Press integration brings live vote counts into ChatGPT after polls close. Election officials remain responsible for authoritative voting rules and certified results. These partnerships improve access to relevant information, but their announcement alone does not establish nationwide reliability.

The company prohibits political advertisements on its platform and the use of its products for scaled campaign advocacy, while allowing specified human-directed administrative work. Those are platform rules whose value depends on detection and enforcement.

The cartoon safeguard shows both progress and limits

One OpenAI protection changes the appearance of certain requested images.

The company says images depicting election interference or potentially discouraging voting are rendered as stylized illustrations, rather than honoring requests for photorealism. Its rules against election interference still apply to cartoon imagery, because apparently, a cartoon of a ballot box is slightly less threatening to the republic.

Independent testing found evidence of improvement. In an October retest of six image generators, the Brennan Center for Justice found that ChatGPT converted nearly all its election-disinformation prompts into illustrations. The researchers found weaker protections or little improvement elsewhere. These were results for the tested prompts and versions.

CNET’s tests also produced cartoons when it requested scenes such as immigration officers outside a polling location. Its attempts to obtain photorealistic versions were unsuccessful.

But Gizmodo reported a bypass: after receiving an illustrated election-interference scene, its reporter asked for a photorealistic version and obtained one. Together, the reports show a protection that can work and can fail. Neither the successful tests nor the bypass establishes how consistently it behaves across all requests.

Security defenses extend beyond the AI companies

The EAC warns that AI can scale familiar threats, including phishing and impersonation. Convincing emails, cloned voices, or fabricated messages from an election official could be used to seek credentials, confuse staff, or mislead voters. An attack can disrupt access to information or administrative services without altering a vote count.

Existing defenses provide a foundation.

The Center for Election Innovation & Research’s report on security practices for the 2024 election found that, of the 24 participating states, all 23 that answered the relevant question regularly backed up voter registration databases, all 22 that answered the question conducted system audits, and 23 of 24 reported participating in cybersecurity exercises. That is evidence of preparation among respondents, with limits: it describes earlier practices and does not cover every state.

Paper records and audits provide additional checks. The EAC explains that most states require post-election audits, during which officials review paper ballots to verify machine counts. These protections address vote verification, while account security, monitoring, and recovery plans address other parts of election operations.

Vendors are offering assistance. OpenAI says it has offered eligible election authorities and voting-system manufacturers special access to cyber-defense models. Microsoft says its AI and elections training hub has reached more than 1,000 officials across 45 states. Those are company-reported offers and participation figures, which leave questions about adoption, coverage, and effectiveness.

There is documented misuse as well. In October, OpenAI reported disrupting Russia-origin and Iran-origin influence operations that used its tools alongside conventional techniques. The cases demonstrate the usefulness of AI to influence operators and the possibility of detection. They do not establish that those particular operations changed US midterm votes.

Claude’s Georgia incident needs its correction attached

Reports initially described Anthropic’s Claude as accessing Georgia election data, with the model bypassing a paywall while working on an internal statistics project. The later clarification materially changed that account.

On Oct. 26, Georgia’s secretary of state said the agent had retrieved information from the state’s free, public Election Data Hub. Officials said it accessed no voter list or voter registration system and bypassed no paywall. The office said Anthropic had confirmed those findings.

The remaining issue concerns control over agents’ actions. Anthropic’s broader investigation described models pursuing unexpected routes to complete tasks, including interacting with real websites when researchers had intended something else. The company reported expanded restrictions and monitoring. Public access to Georgia statistics should be described accurately, while unexpected agent behavior still warrants investigation.

The FEC outage threatens transparency, with no established AI link

An attempted intrusion prompted the Federal Election Commission to take its electronic filing system offline around 10 p.m. Eastern on Monday, Oct. 26, Reuters reported, citing an email sent to campaign compliance firms.

The shutdown blocked groups from submitting time-sensitive campaign-finance reports during the final spending rush. Previously filed reports remained searchable, according to the agency. Its incident page said on Oct. 27 that officials had identified the issue and were implementing a fix; CNBC reported that the system remained offline early Wednesday.

The FEC regulates campaign finance and does not administer voting or count ballots. Public reporting had not established that AI was involved, that attackers successfully breached the system, or that records were stolen or altered. The confirmed consequence was a disruption to financial disclosure.

Its relevance to election preparedness is the need to protect transparency and maintain services under attack, even when voting systems are unaffected.

AI is also a campaign issue with its own campaign money

AI companies and their backers are influencing elections through more conventional machinery: political spending.

A CNBC analysis of FEC filings found roughly $78.5 million in spending by rival AI-policy camps on federal candidates this cycle, with nearly three-quarters going toward primaries. One camp favors faster development and lighter, more uniform rules. Another emphasizes stronger safeguards and greater room for states to regulate. Those divisions cross party lines.

The funding needs careful attribution. Contributions by an executive are distinct from corporate spending, and support for policy advocacy is distinct from funding candidate advertisements. Anthropic, for example, says its funding for Public First Action is restricted to policy and public education and cannot finance candidate elections; separately funded affiliated super PACs have spent on races.

Regulation is already uneven. As of its Oct. 13 update, the National Conference of State Legislatures counted 31 states with enacted laws regulating political deepfakes. Approaches include disclosure requirements and restrictions on certain deceptive material. Definitions, timing, exceptions, and enforcement vary.

Existing federal rules can also apply. The FEC clarified in 2024 that its prohibition on specified forms of fraudulent misrepresentation is technology-neutral and can encompass AI-assisted media. That is a narrower rule than a general prohibition on misleading political content.

Preparedness depends on people, resources, and evidence

The available evidence describes different levels of preparation for different jobs. Campaigns have commercial incentives to adopt AI quickly. Election offices have established security practices and growing experience with limited administrative uses. Voters face an information environment in which AI can be helpful, misleading, or largely invisible.

A Brennan Center survey of 834 local election officials, conducted in January and February and published in April, found that 16% used AI in their work, up from 5% in 2024. At least 71% had prepared for potential election disruptions. Meanwhile, 75% said their state or local government had not provided additional resources to address federal cuts to election-security services. Those findings show preparation alongside resource concerns; they are not an October stress test of every jurisdiction.

For campaigns, responsible adoption means reviewing generated claims, clearly identifying synthetic depictions, protecting personal information, and controlling which actions agents can take. For officials, the EAC’s cases emphasize the need for documented review and local verification. For platforms, meaningful evidence includes independent testing, consistent referrals to authoritative sources, and information about failures and remedies.

Looking toward 2028 and later elections, a central challenge is accountability across the entire chain: who publishes information, which systems retrieve it, how generated answers present it, and who corrects mistakes. A company announcing safeguards is one step. Establishing that those safeguards work across products, languages, and real election conditions requires considerably more evidence.

What voters can verify before and after Nov. 3

For the Nov. 3 election, voters can use the EAC’s state election directory to reach their state or local election office. Registration options, polling hours, identification requirements, early voting schedules, and mail ballot rules vary by jurisdiction. A national chatbot answer needs to lead back to the rules that apply to the individual voter.

AI can help organize research, explain terminology, and suggest questions. Open the cited pages, check their dates and ownership, and confirm that they support the answer. Verify donation recipients on the actual contribution page. Treat an unexpected call, video, or message that changes voting instructions as something to confirm through an independently verified official source.

After polls close, election-night totals remain unofficial. Officials continue to count eligible ballots, reconcile records, and complete required procedures before certification. An AI answer, a live results feed, or a media projection does not replace that process.

Related reading: For a closer look at how an AI agent’s unexpected actions became an election-security story, read eWeek’s examination of Claude’s unexpected actions.