What MIT researchers learned from 16 million election-related AI responses

It’s July 2024. Vice President Kamala Harris just kicked off a blitz run for the White House after a shock switch-up.

Meanwhile, a team of MIT researchers was working to better understand how chatbots perceive this political environment. They fed a dozen leading LLMs 12,000 election-related questions on a nearly daily basis, collecting more than 16 million total responses through the contest in November. Now they are publishing some conclusions from that process.

As the first big US political race to occur since generative AI went mainstream, the 2024 presidential campaigns happened in a media environment in which the average voter was increasingly looking to chatbots for election information.

The authors wanted to study the impact that shift had on the information voters saw, in the same way that previous research has looked at the role of social media or other emerging mediums.

“Whether and how to impart politically fair information has been a sticking point in discussions about radio, print, social media, and now language models,” lead author Sarah Cen, now an assistant professor of engineering and public policy at Carnegie Mellon University, told us in an email.

Trait trades: The authors found that associations between candidates and certain traits shifted over time, potentially in relation to news events. For instance, after Harris took over the campaign from President Joe Biden, his scores for almost every adjective besides “incompetent” dropped. Harris gained some of those lost associations—“charismatic,” “compassionate,” and “strategic”—while Trump gained in “competent” and “trustworthy.”

The researchers note that these moves are not necessarily causal, as there were other factors at play.

Implicit predictions: While researchers encountered an apparent guardrail against LLMs providing direct election predictions, they did find that models could reveal implicit beliefs about the outcome. Through a series of exit poll-related questions, the authors deduced models’ predictions about which candidate’s voters were “more representative of all voters.”

Tailored responses: The researchers found that, to varying degrees, the models’ responses tended to be swayed by users sharing demographic information, such as “I am a Democrat” or “I am Hispanic.”

“These findings indicate that models can be sensitive to steering, which raises important questions about the trade-offs between the abilities of LLMs to be (helpfully) responsive to user queries and direction while also maintaining neutrality with respect to the election,” the authors wrote.

Cen said one of the ways that AI developers might induce models to provide fairer political information is by encouraging more back-and-forth over issues and avoiding personalized responses.

“There is value in allowing for frictions and slowing things down,” Cen said. “Although developers might want LLMs to give a perfectly personalized answer to a political question in one go, it could be better to start with a fairly generic answer and allow the back-and-forth of a conversation with the user to shape the conversation and allow for more understanding, nuance, and depth.”

With AI answers increasingly supplanting media search results both within Google’s search engine and in external chatbots, Chara Podimata, a co-author and MIT Sloan assistant professor, said long-running studies like these should be conducted for every future election.

“Moving forward this research (and the methodology we propose) should become a staple of every election happening in the US,” Podimata said in an email. “We need to know what information these models are giving, how they are calibrating their responses to different users, and what the models actually ‘believe.’ For that, I think election officials and political scientists will be instrumental in informing the design of our survey for future iterations of the method.”