Voters Are Asking AI About Elections — And Getting Different Answers

More and more voters are turning to AI chatbots to learn about candidates and elections. But a new study from the Massachusetts Institute of Technology shows that these chatbots do not always give everyone the same answers. In some cases, what a chatbot says may depend on who is asking — and the information is not always correct or complete.

The MIT study tested several major AI chatbots during Alaska's primary election last month. Researchers asked Claude, an AI made by a company called Anthropic, about a candidate named Dan Sullivan. Alaska had two candidates with that name in the same Senate race, but Claude focused only on one of them — the Republican who was already in office.

The researchers also noticed that Claude gave slightly different answers depending on the political party of the person asking. When a Democrat asked about Senator Sullivan's views on health care, Claude said he had shown 'some flexibility.' When a Republican asked the same question, Claude said he was 'generally aligned with GOP priorities.' The small differences raised concerns about fairness.

The MIT team is now tracking what AI chatbots say about candidates all across the country during the current election season. They run about 19,000 separate questions through the AI systems each time, changing details like the gender, race, location, and political leaning of the person asking. They built a public website called the LLM Election Observatory where anyone can see what the chatbots are saying.

Other groups have found similar problems with AI and elections. A research group called the Institute for Strategic Dialogue tested more than 2,400 questions across six chatbots. They found that nearly 30 percent of the answers were incomplete, unclear, inaccurate, or outdated.

Researchers in other countries have found problems too. A group called Eko reported that AI chatbots gave Brazilian voters false information about candidates and conspiracy theories. Another study found that newer candidates with less information written about them online may be treated less fairly by AI than well-known politicians.

The MIT researchers say it is too early to draw firm conclusions, but they are sure of one thing: voters should not treat AI chatbots as perfectly neutral sources of information. Adam Berinsky, one of the professors leading the project, said clearly, 'Just because a chatbot gives a confident answer does not mean that it is the correct answer.' The team plans to keep tracking AI responses throughout the rest of the election season.