AI chatbots are getting US midterm election answers wrong, researchers find

What researchers found

As Agence France-Presse on MSN reported, researchers say AI chatbots are adding confusion to the information voters see before the US midterm elections on November 3. The problems include misleading voting guidance, mistaken identification of candidates, and repetition of Russian propaganda.

This is the first federal election in which millions of voters are actively turning to AI tools for political information. Chatbots can help explain things and counter conspiracy theories. But several studies show they often pull from unreliable corners of the open web, which can add to an internet already full of election misinformation.

Wrong dates, wrong people

Forum AI, a start-up that tests AI models, examined eight of them. It found that almost 40 percent of their answers to election questions contained at least one factual error.

One example involved Anthropic's Claude. Asked why voters should pick Representative Don Davis of North Carolina, it listed work as a school superintendent among his qualifications. That was a mix-up with someone else.

Questions about voting logistics went worse than questions about candidates. Google's Gemini 3.6 Flash was asked about voting in Pennsylvania and gave November 3 as the last day to ask for a ballot by mail. The real cutoff came a week earlier. By the study's account, a voter who trusted that reply would have applied too late for the ballot to count.

Forum AI also said the chatbots attributed statements to candidates, officials and observers who never made them, hundreds of times. They often relied on weak sources, including Reddit threads. Anthropic did not answer AFP's request for comment. Google said it is focused on helping voters find reliable information and that it has strict quality and policy protections for election topics.

When propaganda gets repeated

A second worry is that chatbots can be steered by foreign influence campaigns. In late October, AFP's fact-checkers debunked a video that was probably altered with AI. It appeared to show a former student of Texas Senate candidate James Talarico accusing him of sexual misconduct. The accusation had no basis and showed signs of a Russian disinformation effort.

NewsGuard, a group that monitors disinformation, then tested AI tools with the claim. Google's AI Overview (the summary box at the top of search results) repeated it. So did Google's Gemini and Microsoft's Copilot. NewsGuard's wider study covered 10 AI tools and warned that users could be fooled by foreign operations and other bad actors.

The picture shifted over time. When AFP ran its own test after the claim had been widely debunked, the tools cited professional fact-checkers. That shows how much a chatbot's answer can change depending on when you ask.

Google said its systems updated quickly to show the latest reporting. Microsoft called it a fast-moving situation with little reliable information at first, and said Copilot later identified the claim as fabricated.

How many people rely on this

The Institute for Strategic Dialogue, based in London, describes this as the first US election in which every major tech platform has AI tools built in. That makes avoiding them harder than it used to be.

A survey by the nonpartisan National Conference on Citizenship found that two in five Americans have used AI tools to look up election information, such as candidates or ballot measures. Trust is limited, though. Only 40 percent trust the tools for logistics like when, where and how to vote. Sixty percent said they trust them only a little, or not at all, to treat candidates without bias.

The better news

The findings are not all negative. The Brennan Center for Justice, a nonpartisan research group, found that leading chatbots pushed back on common conspiracy theories, such as claims that voting machines were rigged. They kept doing so even when users repeatedly challenged them.

The center said the chatbots often made mistakes, yet acted as sympathetic listeners to skeptical users without endorsing the conspiracy theories. It concluded that they could help fight election disinformation.

Taken together, the studies suggest a split. Chatbots can be useful for explaining and for calmly answering doubts, but they are unreliable for specific facts like dates, deadlines and names.