GovChat: Documenting Government GAI Chatbots Through Press Releases and Public Records Requests
Abstract
In local, state, and federal government, officials use generative AI (GAI) chatbots. Yet little is known about how chatbots influence government work. We present the first empirical study of GAI chatbot use in government, drawing on public records requests for GAI chatbot logs. First, we cataloging deployments by 53 federal agencies, 50 states, and 77 local governments. Second, we introduce GovChat, the first corpus of government GAI chatbot logs: 3,216 chat threads (2023–2026) covering 15 state and local jurisdictions across Texas, California, Washington, Georgia, and Minnesota. Third, we show how government chatbot prompts differs from existing public chat corpora: officials use chatbots for policy analysis (n=651, 20.2%), contracting (n=508, 15.8%) and public outreach (n=319, 9%) alongside ordinary office work, like email drafting. These contain risks: factual errors (19\% of generated legal citations are erroneous), misstatements of law, chatbot drafting of public communications, and delegation of value-laden judgment to chatbots.