A multi-agent AI system from Oxford and Stanford automates data journalism end-to-end, transforming raw CSV files into interactive news articles with minimal human intervention.
The system, called Data2Story, deploys seven specialized AI agents working in concert. The lead "Data Journalist Agent" handles the core workflow: it analyzes datasets, generates story angles, conducts web research to verify claims, and produces interactive visualizations with embedded source links. The agents collectively achieve a 93 percent verification rate for factual statements, meaning nearly all claims in generated articles include traceable sources.
Testing shows reader preference splits based on article type. When compared to shorter human-written originals, 74 percent of readers rated the AI output higher. Against carefully crafted long-form journalism, results flattened to a tie. The finding suggests the system excels at rapid, data-driven reporting but struggles against highly polished narrative writing.
The verification rate addresses a core pain point in AI journalism tools. Most large language models generate plausible-sounding claims without evidence. By integrating web search and source attribution into the agent workflow, Data2Story builds verifiability into the generation process rather than bolting it on afterward.
The multi-agent architecture matters here. One agent alone could not reliably execute data analysis, visualization design, fact-checking, and narrative structure simultaneously. Breaking tasks into specialized agents allows each to focus on a single competency. Coordination protocols between agents ensure outputs chain together coherently.
Real-world deployment faces practical hurdles. Newsrooms need training data tailored to their editorial standards. Verification links require live web access and handle changes in URL structure. Interactive graphics demand web development skills to publish properly. The system appears designed for outlets with technical infrastructure rather than small local news operations.
The tie against premium long-form articles reveals where human judgment still dominates. Story selection, narrative pacing