AI chatbots are becoming political advisers. Their answers can mislead voters

As Americans prepared for the 2026 midterm elections, researchers found that popular AI chatbots sometimes gave voters inaccurate or outdated information. In October testing commissioned by Innovating for the Public Good, 13.9% of answers about voting procedures contained outdated or inaccurate details; 5.9% contained errors that could mislead someone about a deadline, voting day or required step. One example reviewed by NPR gave Michigan voters registration deadlines that were later than the real ones. The figures describe answers in the study, rather than the proportion of voters who received or acted on bad advice.

NPR also tested five chatbots with requests for candidate suggestions in New York races. All initially declined to recommend anyone, but Meta AI, Claude, Grok and ChatGPT subsequently produced candidate lists matching stated liberal or conservative preferences after further prompts; Gemini continued to refuse. The companies said they were working to improve election safeguards, while Anthropic questioned whether part of the researchers’ testing reflected the safeguards available to ordinary users. A separate survey found that 25% of US adults reported using AI to learn about major candidates by mid-October. The findings raise questions about how reliably chatbots distinguish factual assistance from voting advice, particularly when voters need current local information.