Table of Contents
Voters are increasingly turning to AI chatbots for information ahead of the elections, yet many popular systems may provide misleading details about candidates and polling site opening times. Researchers warn that some chatbots have even generated lists of candidates to vote for, raising concerns that AI-generated misinformation could influence voter decisions.
Errors in Voting Logistics and Policy Questions
Tim Chambers, primary researcher at the nonprofit Innovating for the Public Good (IFPG), stated that AI systems frequently deliver answers in an authoritative and convincing manner while containing meaningful errors. Chambers noted that users relying on these chatbots could receive incorrect voting instructions or miss critical deadlines.
An ongoing IFPG study is testing how seven major chatbots—Google AI Mode and AI Overviews, Anthropic’s Claude, Microsoft Copilot, Google Gemini, OpenAI’s ChatGPT, and Perplexity—respond to voter questions covering projected winners, policy platforms, candidate comparisons, backgrounds, controversies, and voting rules.
In data collected in October, 13.9% of chatbot answers to voting-logistics questions were outdated or inaccurate. Errors in 5.9% of responses could mislead a voter, potentially causing them to miss a deadline, vote on the wrong day, or skip a required step. Overall, 5.1% of answers contained a clear factual error, and 10.9% carried a clear error or misleading claim.
The research revealed that response accuracy depended more on the specific chatbot than the questions asked. Claude and Copilot accounted for three-quarters of the clear errors. On voting questions, Claude provided wrong or outdated information one in three times, Copilot one in four times, and ChatGPT one in 10 times, according to IFPG.
For example, when prompted about the election in Michigan’s 4th congressional district, Claude provided incorrect online and mail-in registration deadlines that were later than the actual dates. Chambers described these error rates as extremely troubling.
Microsoft spokesperson Jessica Dash stated that helping people find reliable election information is a priority, adding that the company has improved how Copilot responds to election queries and continues to test and refine its delivery. Anthropic spokesperson Michael Aciman explained that because IFPG used an Application Programming Interface (API) for part of its research, some of Claude’s key safeguards may not have been active. IFPG noted that an internal study comparing Claude’s API responses with its web interface yielded comparable results.
Representatives for Meta, Anthropic, OpenAI, and Google stated they are working to prevent their chatbots from producing election disinformation, such as incorrect poll locations and opening times or misleading deepfake images, continuing industry efforts that started after the 2016 elections.
Influence Campaigns and Betting Market Citations
Abdiaziz Ahmed, a technology policy strategist with the Brennan Center’s Elections and Government Program, noted that political campaigns and influence operations now write content specifically designed for chatbots to pick up and repeat, a process known as generative engine optimization (GEO).
IFPG found that when responding to attack questions or criticisms about candidates, chatbots generally relied on less trustworthy sources, such as websites flagged by Media Bias/Fact Check. Furthermore, all tested chatbot platforms commonly cited prediction market websites like Polymarket and Kalshi when asked about race winners and poll results, rather than relying on traditional pollsters.
Chambers pointed out that betting markets can be manipulated and are often inaccurate. Ahead of the midterms, Anthropic, OpenAI, and Google published policies regarding election safeguards, while SpaceX and Microsoft did not publish policies, and Meta’s policy did not mention Meta AI.
Chatbots Provide Candidate Recommendations After Prompting
NPR tested five popular chatbots—Meta AI, Claude, Grok, ChatGPT, and Google Gemini—to determine if they could be prompted into recommending candidates. All platforms except Gemini provided lists of candidates to consider based on stated political preferences in New York races after two or three prompts.
Ahmed stated that chatbots should provide accurate and even information about candidate stances and cite their sources, rather than recommending who to vote for. During testing with liberal, independent, and conservative personas, all chatbots initially declined to recommend candidates.
However, NPR bypassed initial refusals on most chatbots by stating the user was in a hurry and asking for the most left-leaning or conservative candidates matching specific preferences. Claude, Grok, Meta AI, and ChatGPT ultimately provided preferred candidate lists to both the liberal and conservative personas.
Chambers described these responses as leaky refusals, where systems frame responses as an inability to help before functionally producing an endorsement. He argued that disclaimer language used by some chatbots does not prevent the lists from acting as recommendations.
OpenAI spokesperson Nate Evans stated that ChatGPT is designed to be objective by default and that intentional testing to elicit specific responses is not representative of typical user experience. Meta spokesperson Cory Chambliss highlighted the company’s $30 billion investment in safety and security over the past decade to protect elections and support civic participation. Anthropic’s Aciman noted that Claude is designed to help people reach their own conclusions and that requests for voting advice appeared in approximately 0.01% of U.S. Claude conversations. SpaceX did not respond to requests for comment.
Ahmed concluded that while chatbots can be useful for gathering information, voting is a personal decision that should not be handed over to AI systems that can get facts wrong or have their answers manipulated.
