Large language models (LLMs) that underpin AI chatbots have taken on an expanding role in shaping the information environment. Roughly 25 percent of American adults use chatbots daily including for information gathering and synthesis, entertainment, and other uses. LLMs also increasingly influence information flows in more subtle ways. They are now a primary tool for fact-checking and content moderation on social media platforms. Journalists, election officials, civil society organizations and political campaigns increasingly turn to AI for vital tasks including research and content generation.  

While recent surveys show only a small proportion of American voters knowingly turn to AI chatbots for information about politics and elections, this share is likely to grow in the coming years. Approximately 10 percent of Americans report using chatbots to get news and 60 percent report reading AI-generated summaries of search engine results. This means the reliability of the election information they provide is likely to have increasingly significant stakes for voters and election officials alike.  

Prior research from ISD has demonstrated how adversarial actors can influence chatbots’ behavior by targeting topics where authoritative information is sparse, known as data voids. But the ability for chatbots to shape Americans’ trust in the electoral process extends far beyond the possibility of encountering deliberately poisoned data. These systems must also answer questions about local election procedures accurately and completely — rules that differ from state to state, and often from county to county, and are subject to change between election cycles. 

LLMs’ performance on these questions could impact voter confidence in election processes in 2026 and onwards, potentially even affecting voter enfranchisement. To assess their reliability, ISD assessed six leading models against four metrics:  

• Accuracy and timeliness: do LLMs provide clear, complete, and up-to-date responses to prompts about voter eligibility, election procedures, and safeguards?   
• Responsible debunking: how effectively do they handle contested or unverified claims introduced by users?  
• Authoritative sourcing: what types of sources do LLMs cite in responses to election-related prompts, and in what proportions? 
• Language divergence: does response quality change when users prompt in Spanish rather than English? 

ISD tested each model with 15 generic prompts designed to simulate voter questions about the time, place, and manner of elections, as well as five adversarial prompts on past controversies or disputed claims per state. The resulting analysis assesses responses from default consumer models from six developers: Meta (Muse Spark), xAI (Grok 4.3), DeepSeek (V4 Pro), OpenAI (GPT-5.5), Anthropic (Sonnet 4.6), and Google (Gemini 3.5 Flash). In total, 2,400 prompts and responses are included in this dataset (400 per model). 

This analysis includes findings for prompts tailored to 10 state contexts: Arizona, Utah, North Carolina, Ohio, Texas, Pennsylvania, Michigan, Georgia, Colorado, and Minnesota. These states were selected based on factors including recent changes to election processes or eligibility requirements, ongoing litigation or pending legislation with the potential to change these processes or requirements, and a history of election administration controversies.