AI System Beats Human Journalists in Reader Preference Study

What happened

A team of researchers built Data2Story, a system using seven specialized AI agents (Detective, Analyst, Editor, Designer, Programmer, Auditor, Inspector) that turns datasets into complete news articles with zero human input. In a study of 53 readers comparing 18 AI-generated articles against human-written originals from The Economist, The Pudding, and TidyTuesday, the AI version won all five rated categories, with the largest margin in data transparency at +1.49 on a seven-point scale. Overall, 74 percent of readers preferred the agent article.

Why it matters

The system's core strength is machine-verifiability—93 percent of visible statements can be traced to either runnable code or external sources, compared with a 25 percent baseline for human articles. This addresses a real gap: journalists rarely publish analysis code, making their claims harder to fact-check. The Inspector panel shows the exact data or URL backing each claim, letting readers verify or dispute any figure by running the code themselves.

What to watch

The system currently runs on full autopilot with no human-in-the-loop feedback, which is left for future work. It struggles most against handcrafted long-form design (tied with The Pudding's lavishly designed pieces) and cannot replicate human reporting that explains causation or theory—it only shows what the data contains, not why. The code is available at data2story.github.io and GitHub.