Bad actors seeking to misinform potential voters are taking a new approach.
Rather than creating deceptive articles and websites to fool individuals, they’ve turned their attention to manipulating artificial intelligence. By targeting the data models AI systems are trained on, as well as flooding the web with inaccuracies, groups attempting to influence elections are aiming to corrupt the answers provided by chatbots and those increasingly popular AI summaries that appear atop web browser search results.
It’s a process known as algorithmic poisoning.
Washington State University political scientist Thessalia Merivaki has been watching the manipulative high-tech strategy closely.
“While these influence operations, mainly by foreign actors, have not changed election outcomes, they do raise important questions about the scale of the problem.It’s clear that the intent is to influence public opinion during critical events like elections,” said Merivaki, who serves as Director of WSU’s Thomas S. Foley Institute of Public Service and Public Policy.
It may not come as a surprise that foreign influence organizations have pivoted to algorithmic poisoning: according to the Pew Research Center, nearly half of U.S. adults use AI chatbots, with roughly one-in-four using them daily. AI summaries are also ubiquitous, often the first result search engine users see. election officials and researchers are, therefore, very concerned as we are heading into the U.S. midterm elections.
The scale of the problem
Independent election monitoring and fact-checking organizations have already documented attempts to poison election information ecosystems, including in Australia, Moldova, and Hungary. The flooding of false AI-generated content into the open web is a particular concern for Merivaki, given the scale of effort observed so far.
In 2024, at least 3 million auto-generated articles promoting pro-Kremlin content hit the web, translated into an array of languages and aimed at multiple electorates, according to an article Merivaki co-wrote earlier this year for the Center for Democracy and Technology, where she serves as a Non-Resident Fellow. Researchers have also observed significant efforts to edit Wikipedia articles with misleading information, a naked attempt to contaminant a source of information treated by some AI systems as reliable.
The reason behind the shift away from voter-targeted influence efforts and toward AI-targeted ones is straightforward: It’s easier to scale and disseminate the misinformation. The process also takes advantage of consumers’ inherent trust and reliance on information provided by AI chatbots and summaries. Those that would be wise to get a second opinion.
A recent audit from NewsGuard, which monitors information reliability online, found that chatbots spread falsehoods 35% of the time, a rate that nearly doubled in just one year.
At the same time, Merivaki notes that the largest social media platforms have weakened their content moderation protocols particularly relating to misleading political information. Those platforms now rely on community-driven responses to false or misleading posts rather than dedicated factcheckers, potentially in response to criticism about potential censorship.
Ways to safeguard reliable information
AI isn’t going away, and consumers will likely continue to use these tools regardless of concerns about information accuracy. Fortunately, there are ways voters and election officials can combat AI-driven misinformation efforts.
Anyone looking for information on elections, from specifics about how to vote to details about the candidates and ballot measures, should rely on official county and state election office websites. Election officials must also ensure their website have the clearest information available about the mechanics of voting and what is and isn’t allowed in their states.
For example, some states’ election websites might not explicitly acknowledge online voter registration isn’t allowed, but voters may ask a chatbot whether they can register to vote online in their state. This gap seems minor, but can be a risk, because in the absence of a clear response from an authoritative source – the elections office – , the chatbot will generate a response that may contain false or misleading information to satisfy the results it thinks users want.
A 2024 stress-test of how AI chatbots respond to questions about voting with a disability found that 61% of responses had at least one type of insufficiency, with one in every three answers having incorrect information. And it’s not just an issue with specific tools; every model tested by the Center for Democracy and Technology hallucinated at least once in the testing.
While there are sophisticated efforts to misinform voters, Merivaki is optimistic about the growing awareness of these challenges by the election community, and their strategies to respond.
“We are really seeing increased efforts from election officials to put good information out there, and there are more partners willing to help amplify good information.”
She continued, “That’s one of the roles we play at the Foley Institute, helping to inform students and community members with trusted sources of information and building literacy skills, while also supporting election officials with research-drive recommendations on effective voter communication.”
