New AI Pipeline Automates Data Journalism
- •Researchers developed Data2Story, a seven-agent AI system that converts CSV datasets into verifiable interactive articles.
- •The system achieved a 93 percent verifiability rate for claims, compared to a 25 percent baseline for human-written pieces.
- •User testing showed 74 percent of readers preferred agent-authored articles, though humans remain superior at providing narrative context.
Researchers from Oxford and Stanford have introduced "Data Journalist Agent" (Data2Story), a Claude Code skill designed to automate data journalism by transforming CSV files into interactive, verifiable online articles. The system utilizes a "virtual newsroom" of seven specialized agents that perform tasks ranging from web research and data analysis to programming, layout, and audit. Each article features an "Inspector" panel that allows users to trace every claim, chart, or interactive element back to its source, whether that source is an external URL or a runnable code script that reproduces the figure from the underlying dataset.
In comparative testing, 53 readers evaluated 18 article pairs generated by Data2Story against human-written originals from publications including The Economist, The Pudding, and TidyTuesday. Results showed that 74 percent of participants preferred the agent-generated articles, while 25 percent favored the human versions and 2 percent considered them equal. Data2Story outperformed human writing in categories like data transparency, with a +1.49 lead on a seven-point scale. The system achieved a 93 percent verifiability rate for visible statements, significantly higher than the 25 percent baseline observed in human-written news articles, which often lack published analysis code.
Despite its performance in data-heavy tasks, researchers identified areas where human journalists retain a distinct advantage. Humans outperform the system in providing editorial context—explaining the "why" behind trends—as well as in creative data visualization and the synthesis of complex information. For example, while the agent can generate multiple charts for a single subject, human-designed graphics can better integrate annotations and complex layering to convey primary narratives. The researchers concluded that Data2Story functions as a collaborative tool that handles computation and source linking, leaving high-level editorial strategy and narrative perspective to human journalists. The system currently operates on full autopilot, with future development intended to integrate human-in-the-loop feedback mechanisms.