In the run-up to the 2026 midterm elections, the question of how artificial intelligence will interact with American democracy is no longer theoretical. It is already embedded in the machinery of campaigns, voter outreach, and the daily flow of political information. In 2024, many feared AI would unleash a tidal wave of deepfakes and viral disinformation that would overwhelm the election. That did not happen—at least not in the way people predicted. Instead, AI slipped into the process more quietly, often invisibly, shaping what voters see, read, and trust. President Trump regularly shares AI-generated images on social media, AI-generated attack ads have become routine, and around the world, dozens of elections have now been fought in the presence of these tools. Yet the deeper question remains: not whether AI will cause a spectacular crisis, but how it is quietly rewiring the information ecosystem that voters rely on. To understand this shift, I spoke with Thessalia Merivaki, a political scientist at Washington State University’s Foley Institute. She has spent years studying how the digital information environment affects voter behavior and confidence in election integrity. Her team built a tool called the Election Officials Communications Tracker, which monitors social media posts from election offices across the country. Our conversation centered on a few pressing issues: what happens when more voters than ever get election information from AI chatbots, whether election officials can make authoritative information cut through the noise, and what a more resilient electorate might look like in the age of AI.
The first major shift is that AI has changed how voters search for information. Instead of typing a query into Google and seeing a list of links to evaluate, voters now often receive a single, AI-generated summary that looks authoritative but offers little visibility into where it came from or whether it is reliable. Even when links are included, very few people click through to check the source. This is a fundamental change in the relationship between people and information. In the past, voters exercised judgment by choosing which sources to visit, which headlines to trust, and which sites seemed credible. Now, the choice is made for them by an algorithm that compresses vast amounts of content into a neat answer. This matters enormously for elections because voting procedures are highly localized and detail-oriented. Questions like “Can I vote by mail in my county?” or “What ID do I need to bring to the polls?” depend on where a person lives. AI systems do not always have good, specific, authoritative material to draw on for these questions, so they often surface low-quality or incomplete information. At the same time, election officials, campaigns, and political organizations are also exploring AI tools—drafting outreach content, managing social media, and even scaling entire communications operations, especially in jurisdictions that lack dedicated communications staff. So AI is not just a problem for voters; it is also becoming a tool for the very institutions trying to build trust.
Merivaki argues that AI’s most consequential impact is not the fake content people see directly, but something more structural: algorithmic manipulation, or “poisoning.” This is the effort to corrupt the infrastructure that feeds AI models, so that when voters use chatbots or AI-powered search, the results reflect deliberately polluted information. We saw this in 2025 in elections in Australia and Moldova, where Russian influence networks flooded the internet with content designed not for human eyes, but for AI crawlers to collect and absorb into their training data. That poisoned material then gets fed back to users as authoritative answers. This reveals a serious vulnerability. In the United States, there are more than 10,000 local election jurisdictions. While federal and state policies provide some uniformity, the way information is communicated to voters varies wildly at the local level. The breadth of official election information across the country is broad, but the depth of official content on any single jurisdiction is often very thin. This creates what researchers call data voids—gaps in credible coverage that bad actors can fill with low-quality information. Even without coordinated attacks, these data voids explain why AI outputs about voting procedures are often unreliable. The systems simply lack enough authoritative material to draw from. The result is an asymmetry: official, accurate information from election officials struggles to dominate AI retrieval environments, while voters are left uncertain about what to trust. This uncertainty feeds directly into declining confidence in elections, because people sense that the information they are getting is incomplete or untrustworthy, even if they cannot articulate exactly why.
Merivaki’s research offers a closer look at what election officials are actually doing to fight back. Her team’s Election Officials Communications Tracker collects data on social media communications from state and local election officials, and the findings paint a mixed picture. Social media has become a core communications channel for election officials, and for good reason: it is cost-effective and reaches a large number of voters. But usage is highly uneven. Every state has an official elections account on mainstream platforms, but local election officials’ presence is spotty and mostly concentrated on Facebook. This unevenness matters because a thinner presence means less authoritative and high-quality information circulating in the spaces where voters increasingly spend their time. Still, the research shows that when election officials do communicate on social media, they are not only providing practical information like how to register or how to vote by mail. They are also engaging in trust-building. They explain how elections are kept secure, what safeguards are embedded in the process, who to consider an authoritative source, and even what misinformation and AI are. The data suggests this work pays off. Merivaki found that explicit, focused messaging about how elections are run and the security measures integrated into every step of the process is associated with higher voter confidence—not only among the broader electorate, but also among people who have expressed skepticism in election outcomes. In other words, these communications build trust, even among those who have serious doubts.
But trust-building campaigns are not a cure-all. Merivaki is careful to say that while these interventions are effective in experimental settings, real-world conditions are messier. The effects are real, but they are not long-term, because too many other factors shape how voters think about elections and candidates. One of the most prominent examples of a trust-building campaign is the trusted info campaign created by the National Association of Secretaries of State in 2020. Its core messages have been adopted by many states. The first message is that election officials are the authoritative source of information about elections. The second is that officials are following rules and procedures to ensure elections are secure. The third is that elections are run by humans—trained professionals who are part of the community. Merivaki’s research finds a strong relationship between repeating the message that election officials are the source and increased voter confidence. The more this message is communicated, the more voters name their local election official as a top source of information. In 2024, a quarter of voters said their primary source of election information was their election official. That is not a majority, but it is a meaningful number, and it shows progress. However, these campaigns have a critical weakness: timing. They are most effective in the pre-election phase. The hardest period is after voters cast their ballots, especially after election night, when winners are declared. This is when the “loser effect” takes hold—the bitterness of defeat, especially across partisan lines. Add to that delays in counting, inconsistent rules across states, and candidates themselves claiming the election is stolen, and you create conditions for significant declines in trust. Once that trust is lost, it is very difficult to restore, because the decline is grounded in factors that have little to do with how elections are actually run. Trust-building campaigns cannot easily address those dynamics.
So what can be done? Merivaki argues that the old mindset of rapid response, debunking, and fact-checking is not enough. Voters need more than corrections; they need familiarity with the rules and procedures of elections. When voters understand how the process works, they are better equipped to resist misinformation. If a candidate claims there is massive fraud with mail voting, a voter who has read about the safeguards and knows how the process works is more likely to think, “That doesn’t match what I know.” This kind of resilience comes from education and exposure, not just from reactive fact-checking. On governance, Merivaki emphasizes that any regulatory framework must start from the reality of local election officials. They are not tech companies. They do not have engineers on staff, and many jurisdictions do not even have a dedicated communications director or IT person. Yet they are expected to serve all these roles with limited capacity and funding. This context must be the starting point for any AI governance framework that is actually going to work. Operationally, the most critical requirement is human-in-the-loop oversight at every stage where AI interacts with public-facing output. Even routine AI-assisted communications must be reviewed by a human. This check is the mechanism that maintains institutional legitimacy, because in elections, every service is ultimately run by humans. If AI is allowed to operate without human oversight, the reputational and trust costs could be significant. Merivaki also calls for airtight, documented prompt libraries for election officials to use, protocols for vendors, more transparency in how AI models are built and audited, and clear rules about who approves the output. She emphasizes the need for chain-of-custody features as AI tools are integrated. On data and privacy, officials need to be explicit about what is considered sensitive information and how those rules apply to internal communications. These are not just technical details; they are essential safeguards for protecting the integrity of elections and the trust voters place in them. As we move toward 2026, the challenge is not simply to prevent AI from causing chaos, but to build a system where authoritative, human-verified information can still rise above the noise.