As artificial intelligence chatbots become an increasingly common source for voters seeking election information, concerns about the accuracy and impartiality of their responses have surfaced. A recent research initiative led by the Massachusetts Institute of Technology (MIT) highlights how AI-powered language models deliver varied and sometimes incomplete information about political candidates and issues, depending on user identity and other factors.
The MIT project, publicly unveiled this week, analyzed chatbot responses ahead of recent primary elections, including the Alaska Senate race. In one instance, an AI model named Claude, developed by Anthropic, was asked about “Dan Sullivan’s position on health care.” Instead of addressing both candidates named Dan Sullivan, the model focused exclusively on the incumbent Republican senator. Furthermore, when the question was posed from a Democratic perspective, Claude described the senator as having “shown some flexibility” on health care, while a Republican query resulted in a response that the senator was “generally aligned with GOP priorities.”
This experiment is part of a broader effort by the MIT team, which includes political science professors Adam Berinsky and Charles Stewart III along with assistant professor Chara Podimata, to systematically track how large language models respond to election-related queries across various demographic and political identities. The researchers have been running automated sweeps comprising 19,000 different query combinations to assess how models tailor answers based on user characteristics such as gender, race, location, and political affiliation.
The study builds on prior research from the 2024 presidential election cycle, which found that AI models not only adapt to evolving political campaigns but also internalize biases, portraying candidates with subjective tones that imply compassion or combativeness. With the midterm elections underway, the team has developed a public dashboard called the LLM Election Observatory to monitor AI behavior in real-time. While the researchers caution against drawing firm conclusions about systemic bias or accuracy at this stage, they emphasize the need for vigilance, noting that confident chatbot replies do not guarantee correctness.
Other organizations have reported similar concerns. An analysis by the Institute for Strategic Dialogue examined thousands of chatbot responses in June and found nearly 30 percent of answers in English were incomplete, unclear, inaccurate, or outdated. Errors included incorrect election dates and outdated rules. The Brennan Center for Justice reported that chatbots resisted spreading false election conspiracy claims but occasionally generated misleading images and videos related to voting. Additionally, research groups have identified that lesser-known candidates may be disadvantaged by AI-generated electoral information due to sparse online data.
Anthropic, the developer of Claude, stated that their model is trained to treat differing political views impartially and undergoes bias testing prior to each launch. OpenAI did not respond to requests for comment. Meanwhile, the MIT study found that the same question about Texas Democratic Senate candidate James Talarico elicited different answers depending on the chatbot asked and the political identity the chatbot believed was posing the question.
As AI continues to reshape how voters access political information, researchers underscore the importance of understanding how algorithmic tools influence democratic processes. The MIT team aims for their observatory to provide ongoing, transparent insights into how AI interprets and potentially distorts election-related information.