On any given election day, older voters turn out in far greater numbers than younger ones — and are far more likely to identify with a political party.
Semra Sevi wondered whether artificial intelligence might help close the gap, by giving young, unaffiliated voters clearer information about where the parties stand on the issues they care about most.
To investigate this question, she and colleagues designed a series of three experiments in which nearly 2,900 politically unaffiliated voters under 35 were invited to engage with an interactive Voting Advice Application (VAA) chatbot — one that didn’t just answer questions, but held a back-and-forth conversation tailored to each user’s own policy priorities.
We found big effects on knowledge gain, though not an effect on taking up a party as a result of this knowledge.
Their findings were published in a PNAS article this past December. The results are striking, though not entirely in the way one might expect.
“We found big effects on knowledge gain, though not an effect on taking up a party as a result of this knowledge,” says Sevi, assistant professor in the Department of Political Science in the Faculty of Arts & Science. “There was about a 13-percentage-point improvement in terms of understanding where parties stand, especially on issues the subjects personally cared about. It also produced some smaller improvement in understanding other issues voters hadn’t flagged as a core concern, about 4 percentage points. But we found little effect on party preference or voting intention.”
Encouraged by the chatbot’s ability to educate, Sevi plans to deploy the tool with different groups in the future.
“We want to give people clear personalized information about party positions and see if they can understand politics better as a result,” she says. “Will they feel closer to a party, and will it change who they might vote for?”
VAA are not new — they’ve been used around the world for years. In Canada, the best known is Vote Compass, a voting advice tool used by media outlets during election campaigns.
But Sevi says many applications can be static and off-putting, taking the form of lengthy questionnaires. They also tend to attract voters who are already politically informed, the very people who need them least.
“Our VAA bot talks to users conversationally; it’s an interactive way for young people to learn about politics. And given that young people like gamification — where they can get fast replies to personalized questions — we’ve designed a VAA bot that meets them where they’re currently at,” she says.
The third and largest of the three studies, which involved 2,000 participants, was conducted close to the 2024 American federal election, a period when political interest tends to peak. She says the broader political climate in the U.S. may help explain why knowledge gains didn’t translate into stronger party attachments.
“Even though these young people did form party knowledge, they may not have wanted to align with a party because of how polarizing politics are in the U.S. right now,” Sevi says. “In the Canadian context, we don’t have the same sort of dynamic. So, it’s possible that when we repeat the study here, we might see an uptake in party identification.”
One of the general criticisms of AI bots is that they can sometimes “hallucinate,” producing false or misleading information. Sevi’s team took pains to ensure accuracy by grounding the chatbot’s responses in verified sources.
There’s a long literature that says people vote for the party, not the issues. But what we’re finding is that issues do matter.
“We’re avoiding misinformation by using what’s called the retrieval-augmented generation method, which grounds the answers in real sources: party platforms, speeches and voter guides.”
The idea began as a postdoctoral research project Sevi undertook at Columbia University during the pandemic, working alongside political scientist Donald Green. In those early days, the team reached out to young voters through videos watched over Zoom.
But as large language models began rapidly transforming what was possible in human-AI interaction, the team saw an opportunity to harness these new tools for civic education. When fellow Columbia political scientist Yamil Velez joined the project, he brought the technical expertise needed to build the VAA bot, making a far more scalable and engaging methodology possible.
“For follow-up studies, we’re going to deploy the VAA bot in Canadian high schools,” she says. “We’ll be talking to teachers to see what kinds of challenges they face teaching young people about civics, political parties and where they stand. Then we’ll tailor the VAA bot that we tested in the American case."
She also plans to work with newly arrived immigrant groups who, like youth, may be seeking political information.
“However, with this chatbot, we don’t persuade at all: the design is neutral on purpose,” Sevi adds. “That’s important for us because when we work with teachers in Canada, we don’t want anyone to think that we’re working with one party over another.”
So far, Sevi’s work shows that when carefully deployed, AI can play a crucial role in the political education of unaligned voters, even if changing minds takes more than better information alone.
“There’s a long literature that says people vote for the party, not the issues. But what we’re finding is that issues do matter,” she says.
“We see a positive outcome from this experiment, because knowledge gained is important — even if young people don’t align with a party as a result.”