Published • loading... • Updated
Nearly a third of AI chatbot replies to voting questions inaccurate, outdated: Study
Researchers found 29% of chatbot answers were incomplete, inaccurate or outdated, with Spanish prompts producing worse results across all six models.
- The Institute for Strategic Dialogue tested six leading AI chatbots—Meta , xAI , DeepSeek , OpenAI , Anthropic , and Google —using 2,400 prompts across 10 states to assess voting information accuracy.
- Roughly 25 percent of American adults use chatbots daily for information gathering, and approximately 10 percent seek news from them, while voter reliance on AI for election details is expected to grow in coming years despite current low awareness.
- Each model faced 15 generic voter prompts plus five adversarial prompts per state, evaluated on accuracy, handling of contested claims, source authority, and Spanish-language performance. The 10 states were selected based on recent election process changes, ongoing litigation, and documented controversies.
- Local election procedures differ by state and county and change between cycles, creating complexity that AI systems must navigate accurately as journalists, election officials, and campaigns increasingly rely on chatbots for election-related research and content generation.
- As voter adoption of chatbots accelerates while these systems struggle with complex, localized election procedures, prior research showing adversarial actors exploit data voids underscores urgent need for AI reliability standards in the coming years.
Insights by Ground AI
8 Articles
8 Articles
AI Chatbots Flub Basic Voting Questions
“Leading AI chatbots provided inaccurate, unclear or outdated replies on 29% of basic questions about elections and voting procedures in tests this year ahead of the midterm elections, according to a new report,” NBC News reports.
·New York, United States
Read Full Article
Coverage Details
Total News Sources8
Leaning Left2Leaning Right0Center6Last UpdatedBias Distribution75% Center
Bias Distribution
- 75% of the sources are Center
75% Center
L 25%
C 75%
Factuality
To view factuality data please Upgrade to Premium
