Before the Ballot Box, a Ballot Bot?

Amid a slate of unexpectedly competitive races, a surge in early voting and rising concerns across the political spectrum about election interference and ballot tampering, crunch time at election offices nationwide during this midterm election is crunchier than ever. 

“People are calling. People are emailing. State legislators and county council members are forwarding questions and worries that they get,” said Alysoun McLaughlin, a former election official and head of the Election Resilience Lab at the University of Maryland. 

A team led by Cody Buntain and Giovanni Luca Ciampaglia, assistant professors in UMD’s College of Information, hopes to help election offices by developing an artificial intelligence-powered tool to answer incoming questions quickly, accurately and in an engaging way, ultimately conveying a reassuring truth: U.S. elections are overwhelmingly safe and well-run.

With that established in the minds of voters, the election workers — whom a recent UMD study co-authored by McLaughlin found were underpaid and overworked — could focus more energy on other urgent tasks. 

The research team’s custom AI chatbot named VIOLETS (Voter Information from Official Local Election Trusted Sources) was trained on vetted information from the Maryland State Board of Elections and the Montgomery County Board of Elections offices.  

“Any answer that we provide needs to be citable back to a piece of information from the state or county websites, or their documentation, their manuals,” Buntain explained.

Although state and local governments share reliable, detailed information about election processes on their websites and in press releases, not everyone reads these sources, Buntain said. Instead, voters are increasingly turning to AI tools to answer their election-related questions: 1 in 5 this season, according to polling released this month by the American Association of Political Consultants.

“People want easy information,” said research team member Do Won Kim, a UMD doctoral candidate in computational social science, and asking a chatbot is often simpler than sifting through pages of information on a website. 

“It's more efficient. It's more engaging from the beginning,” Kim explained. “The chatbot prompts you back, it answers you.” 

These answers don’t always get it right, though. Election procedures vary significantly across the country, and generally available AI tools can pull information from unofficial sources or may misapply another state’s election rules in their answers. It’s an outgrowth of the shift away from local news — often featuring journalists with decades of political knowledge about specific cities or counties — to a more consolidated, national news environment. 

“The time is behind us when most voters would be getting their news from a local reporter,” McLaughlin said. 

When an election procedure gets picked up by national media outlets, local officials often get flooded with inquiries about a policy that may not pertain to their voters, she said. This year she has seen buzz on social media warning voters not to allow an election worker to make any marks on a ballot, implying that this could be a nefarious attempt to invalidate votes. “Well, in Maryland, the poll workers do not make marks on ballots, but in North Carolina, they do,” she said. 

“The rules really are very Balkanized,” she added. “The strength of our election system is in its decentralization, but that also makes it more difficult for voters to follow what's going on when they're consuming national resources.”

Even the AI chatbots that some local governments are using to extend the offerings on their 311 information lines don’t always capture the full nuance of election procedures correctly. The VIOLETS developers hope to fill these gaps. 

The experiment, which is running now, is testing how voters who get information by interacting with VIOLETS compare to those who access the same information by reading official websites. Two separate groups will access the curated information either through a back and forth with the chatbot or by using a modified search engine that can retrieve this verified information. 

Before and after measurements of political knowledge, perceptions of the integrity of local elections, and other indicators will reveal if the interactive chatbot exchange is more effective than simple searching and reading. The team suspects it will drive higher levels of knowledge, trust and engagement, while lowering the influence of misinformation about election procedures.

At Maryland Day on campus earlier this year, the team conducted a trial of the chatbot, which helped the research team get insights about key guardrails built into the bot. For example, Buntain said, “it won't answer questions about particular candidates. It won't answer questions about their platforms.” 

The team hopes to roll out the tool in future elections and has been in talks with initiatives at the state level to include it in the Go Vote Maryland app, perhaps in time for the 2028 presidential election. 

“We can throw lots of resources at trying to produce more information in more forms,” McLaughlin said. “But really learning how voters are experiencing that information: What is most effective in meeting them where they are? What increases their engagement? What reduces their engagement? These are all the kinds of questions that I'm excited to see come out of the VIOLETS research.”