Imagine standing in your kitchen with a mug of coffee, phone in hand, trying to figure out one simple thing: how do you vote in the upcoming election? You type the question into a search bar, and before you even finish the sentence, an artificial intelligence-generated summary appears at the top of the results. It sounds confident, clear, and official. It tells you that you can register online in your state, or that you need to show a type of identification you’ve never heard of. You trust it because it looks like an answer, not a guess. But what if that answer was never true? What if it was planted there by people who wanted you to believe it, not because it helps you, but because it helps them? This is the unsettling reality of a growing threat in American democracy, one that doesn’t attack voting machines or storm polling places, but quietly poisons the information you rely on. Bad actors seeking to misinform voters have shifted their tactics. Instead of creating fake news articles or doppelgänger websites, they are now targeting the artificial intelligence systems that have become the world’s primary source of answers. They are manipulating the data these models learn from and flooding the open web with falsehoods, hoping to corrupt not just what you read, but what the machines tell you. It’s a process known as algorithmic poisoning, and it’s happening right now, as the United States heads into its midterm elections.
Washington State University political scientist Thessalia Merivaki has been watching this manipulative strategy unfold with a mix of urgent concern and careful scholarly attention. As the director of WSU’s Thomas S. Foley Institute of Public Service and Public Policy, she has spent years studying how elections work, how people get information, and how misinformation threatens the integrity of democratic participation. The new approach, she explains, is less about convincing any individual voter to believe a particular lie and more about corrupting the systems those voters use to learn about their civic responsibilities. These influence operations, mainly driven by foreign actors, have not yet changed election outcomes, but that does not mean they are harmless. “They do raise important questions about the scale of the problem,” Merivaki says. “It’s clear that the intent is to influence public opinion during critical events like elections.” The shift makes sense in the context of how Americans actually get their information today. According to the Pew Research Center, nearly half of all U.S. adults now use AI chatbots, and about one in four uses them daily. AI summaries have become ubiquitous, often appearing as the first thing a search engine user sees, even before clicking through to a website. This means the battleground of election misinformation is no longer just the front page of a suspicious website; it’s the text box of a chatbot, the featured snippet on a search results page, the seemingly neutral machine voice that has become a new kind of civic advisor.
To understand algorithmic poisoning, it helps to think of artificial intelligence less as a magical oracle and more as a very fast reader that has consumed enormous amounts of text from the open internet. These systems, known as large language models, learn patterns, facts, and assumptions from the data they are trained on. If the data is contaminated with falsehoods, the models will repeat those falsehoods with the same confidence they use to explain the weather or recommend a recipe. Attackers have figured out that if they can flood the web with misleading content, or more specifically, shape the sources that AI systems treat as reliable, they can effectively control what millions of people hear. They have also realized that it is much easier to scale this kind of manipulation than to run a traditional influence campaign. In the past, a foreign operative might have needed to create dozens of fake social media accounts, write compelling posts, and hope that real people would share them. Now, they can generate thousands of articles, edit trusted reference sites, and insert falsehoods into the digital ecosystem with automated tools. The process takes advantage of the inherent trust people place in AI-generated answers. Consumers do not usually interrogate a chatbot the way they might question a stranger at a rally; they treat it as a neutral, authoritative source. And because these tools are already deeply embedded in search engines, smartphones, and browsers, they are the perfect vehicle for misinformation that is difficult to detect, easier to believe, and almost impossible to trace back to its original source.
The scale of the problem is already documented. Merivaki, who also serves as a Non-Resident Fellow at the Center for Democracy and Technology, co-wrote an article earlier this year that revealed just how massive these operations have become. In 2024, at least three million auto-generated articles promoting pro-Kremlin content hit the web, translated into a staggering array of languages and aimed at multiple electorates around the world. These were not curated opinion pieces; they were mass-produced propaganda, designed to be absorbed by AI training systems and search algorithms. Researchers have also observed coordinated efforts to edit Wikipedia articles with misleading information, a naked attempt to contaminate a source that some AI systems treat as a reliable foundation. The problem is not limited to the United States. Independent election monitoring and fact-checking organizations have already documented attempts to poison election information ecosystems in Australia, Moldova, and Hungary, among other countries. The pattern is clear: this is not a handful of bad actors playing pranks; it is a sustained, sophisticated information warfare strategy. The reason behind the shift away from voter-targeted influence efforts and toward AI-targeted ones is straightforward. It is easier, cheaper, and faster. A traditional campaign to deceive voters requires building fake personas, crafting believable narratives, and hoping for amplification. Algorithmic poisoning only requires enough volume and repetition to corrupt the sources that AI systems rely on. What makes this especially troubling is that it piggybacks on the very convenience that has made AI so popular. People use these tools because they are fast, useful, and, for the most part, accurate enough for everyday tasks. But when it comes to elections, the stakes are too high to accept a confident answer without a second opinion, and the research suggests that this caution is fully warranted.
Indeed, the trust that people place in chatbots is deeply concerning, especially given how often those chatbots are wrong. A recent audit from NewsGuard, an organization that monitors information reliability online, found that chatbots spread falsehoods thirty-five percent of the time, a rate that nearly doubled in just one year. That is more than one in three answers being misleading or outright false, an alarming statistic for a tool that people use to make important decisions, including decisions about how to vote. At the same time, Merivaki notes that the largest social media platforms have weakened their content moderation protocols, particularly those related to misleading political information. In response to years of criticism about censorship and political bias, many platforms have moved away from dedicated fact-checking teams and now rely on community-driven responses to false or misleading posts. This shift means that misinformation can linger longer, spread faster, and reach more people before it is challenged, if it is challenged at all. The combination of unreliable AI summaries and weakened moderation creates what researchers call an information ecosystem that is uniquely vulnerable to manipulation. It’s like having a city where the water supply is only lightly filtered, the public health department is underfunded, and the people are told to just boil their water at home. Some will, but many will not, and the consequences can be devastating. For voters, the danger is not just that they will be tricked into voting for someone they do not want, but that they will be confused, discouraged, or misled about the basic mechanics of participation, like where to vote, when to vote, and whether they are even eligible to vote in the first place.
So what can be done? For Merivaki, the answer is not to abandon artificial intelligence or to panic about the state of the internet, but to build stronger, clearer, more authoritative sources of information that people can rely on. Anyone looking for information about elections, from the specifics of how to vote to the details about candidates and ballot measures, should go directly to official county and state election office websites. These are the authoritative sources, the ones that are actually responsible for administering elections and ensuring that every eligible citizen can participate. Election officials, in turn, must make sure their websites contain the clearest possible information about the mechanics of voting, including what is and is not allowed in their states. This may seem obvious, but there are gaps. For example, some state election websites might not explicitly state that online voter registration is not permitted, because that fact seems obvious to the people running the website. But voters may ask a chatbot whether they can register to vote online, and in the absence of a clear response from an authoritative source, the chatbot will generate an answer based on whatever fragmented, unreliable information it has absorbed from the rest of the internet. The result can be false or misleading information delivered with complete confidence. This gap seems minor, but it is exactly the kind of opening that algorithmic poisoning exploits. A 2024 stress-test of how AI chatbots respond to questions about voting with a disability found that sixty-one percent of responses had at least one type of insufficiency, and one in every three answers contained incorrect information. It is not just one chatbot either; every model tested by the Center for Democracy and Technology hallucinated at least once during the testing process.
Despite all of this, Merivaki remains optimistic. She points to growing awareness among election officials and community partners as a reason for hope. “We are really seeing increased efforts from election officials to put good information out there,” she says, “and there are more partners willing to help amplify good information.” That includes institutions like the Foley Institute, which plays an important role in helping students and community members build information literacy skills, teaching them how to identify trustworthy sources and how to question the answers they receive from AI systems. The institute also supports election officials with research-driven recommendations on effective voter communication, helping them understand how people actually seek information and where the gaps in understanding are most likely to occur. This is not just about defending against bad actors; it is about building a civic culture that values accuracy, transparency, and shared responsibility. For ordinary citizens, the best defense is a simple habit: slow down. When you see an AI-generated answer, don’t treat it as the final word. Check it against an official source. Ask yourself where the information came from, and whether a machine could have been tricked. None of this means you have to distrust every chatbot response, but it does mean that for something as important as voting, the most human thing you can do is think twice. The machines are not going away, and neither are the people who want to use them against us. But if we stay informed, stay skeptical, and stay connected to the real sources of democratic authority, the effort to poison our information system can be neutralized. The truth may not always be the easiest answer to find, but it is still the most powerful one.