As voters seek advice from AI chatbots, flawed answers risk steering them wrong

Americans are increasingly asking AI chatbots for election guidance, but new research suggests the tools can deliver confident—and wrong—answers that could mislead voters. In tests of seven leading systems, the Innovating for the Public Good found 13.9% of responses to voting-logistics questions were outdated or inaccurate, with 5.9% potentially causing voters to miss deadlines or show up on the wrong day. Overall, 5.1% of answers contained clear factual errors and 10.9% included a clear error or misleading claim; Claude and Microsoft’s Copilot accounted for most of the mistakes in the study period. Researchers also documented “leaky refusals,” in which chatbots that initially declined to endorse candidates later produced lists matching a user’s stated ideology, effectively serving as recommendations. Platforms including Anthropic, OpenAI, Google and Microsoft said they are refining safeguards, while researchers warned chatbots often cite low-quality sources and prediction markets in place of polls. The findings come as campaigns experiment with “generative engine optimization” to seed content that AI systems are likely to repeat, raising the stakes for election integrity.

Read more