The CitizenQuery Benchmark: A Novel Dataset and Evaluation Pipeline for Measuring LLM Performance in Citizen Query Tasks

High-stakes misinformation risks require careful evaluation: The paper emphasizes that LLM errors in citizen queries about government services—covering benefits, taxes, immigration, and health—could have "severe, negative, likely invisible ramifications" for individuals relying on the information, making rigorous benchmarking essential before deployment in public sector contexts.

Current models show inconsistent performance across the board: The study found that while 11 tested models have "distinct performance profiles across model families," each remains "competitive," but critically, "high variance undermines utility"—meaning unpredictable quality makes these systems unreliable for consistent citizen guidance.

Models avoid admitting uncertainty while over-explaining: The benchmark revealed that "abstention is low and verbosity is high," indicating that LLMs tend to provide lengthy answers even when uncertain rather than acknowledging the limits of their knowledge—a concerning pattern when citizens need trustworthy, honest guidance on important matters.

Synthetic benchmarks from real government data enable systematic testing: The CitizenQuery-UK dataset of 22,000 query-response pairs derived from gov.uk information provides a structured methodology for evaluating LLM factuality, abstention rates, and verbosity, offering a replicable approach for assessing AI systems before public deployment.

Trustworthiness requires transparency and accountability: The paper's conclusion that "more trustworthy AI requires acknowledged" limitations suggests that responsible deployment in citizen-facing government services demands explicit acknowledgment of model uncertainties and constraints, not just technical performance metrics.