Expert benchmark measures AI news quality in sourcing and fairness
NewsBench: Expert-Grounded Evaluation of Epistemic Quality in AI News Reporting
Artificial Intelligence
Summary
As AI increasingly provides news and political information, it’s important to check how accurate, unbiased, and reliable it is. The authors created NewsBench, a new way to test AI news outputs based on standards made by experienced journalists and analysts. Experts first set detailed rules and labeled thousands of examples, then AI tools were trained to judge these qualities so testing could scale. This system helps people see how well different AI models do when reporting facts, naming sources, and staying neutral.
What this means in practice
- •For newsroom teams: Assess accuracy and fairness of AI-generated news content to maintain editorial standards.
- •For social media platforms: Use automated checks to flag or evaluate AI news content quality to improve information reliability.
Authors
Matt Wilde, Jillian Fisher, Andrew B. Hall, Kathryn Salam, Robbie Goldfarb
Abstract
As frontier AI systems increasingly serve as sources of political and current-events information, evaluating the epistemic quality of their responses is now especially consequential. Existing evaluations often focus narrowly on ideological alignment, rely on measures not validated for this setting, or depend on reference labels derived from public opinion or non-expert annotation, potentially producing misleading or miscalibrated results. To address these limitations, we introduce NewsBench, an expert-grounded benchmark that reframes evaluation around editorial standards developed by senior practitioners in journalism, policy, and intelligence analysis. NewsBench operationalizes quality across three dimensions, source quality, factuality, and neutrality, using a two-phase pipeline in which domain experts first define and calibrate evaluation guidelines, which are then transferred to automated AI judges validated against expert-adjudicated gold labels. The benchmark includes approximately $3,000$ evaluation prompts, with roughly $500$ refreshed monthly to capture emerging current events, roughly $2,500$ human expert-annotated items for calibration and validation, and a suite of AI judges designed to enable scalable evaluation. To validate our methodology, we report both inter-rater reliability among human experts and AI judge--human agreement across all three dimensions. Finally, we apply NewsBench to four frontier models, revealing distinct and non-overlapping failure profiles in sourcing, factuality, and neutrality.