A new AI tool called Data2Story, developed by researchers from Oxford and Stanford, automates the creation of interactive news articles from CSV files while ensuring all claims are verifiable. The system uses seven AI agents to process data, generate content, and ensure transparency. The tool's core feature, the 'Inspector' panel, links every statement and visual element to its source, allowing users to trace claims back to code, data, or external URLs. This ensures that all visible statements in the output are traceable, with 93 percent of claims linked to evidence. The system is designed to handle datasets ranging from World Cup schedules to ArXiv trends, offering a structured approach to data journalism. | Image: Lin et al.
Data2Story's workflow is built around a 'virtual newsroom' of seven specialized agents, each handling a distinct task in the editorial process. The 'Detective' agent conducts web searches for context, while the 'Analyst' runs code to analyze data instead of guessing numbers. The 'Editor' selects which findings drive the narrative, and the 'Designer' chooses the appropriate medium for presenting the information. The 'Programmer' constructs the HTML page, the 'Auditor' checks for layout errors, and the 'Inspector' ensures all content is traceable to its source. This multi-agent approach aims to streamline data journalism without sacrificing transparency or accuracy. | Image: Lin et al.
The researchers tested Data2Story against human-written articles from sources like The Economist, The Pudding, and TidyTuesday, with 53 readers rating both versions across five categories. Data2Story outperformed human-written articles in transparency, with a score of +1.49 on a seven-point scale. Overall, 74 percent of readers preferred the agent-generated articles, while 25 percent favored human originals. However, the system still struggles with creative design and dense visualizations, areas where human journalists excel. | Image: Lin et al.
Source: thedecoder
