Data2Story: 7 AI Agents That Out-Verify Human Data Journalism

Oxford and Stanford researchers built a seven-agent system that takes a CSV file and produces a complete, interactive news article — and it out-verifies human-written data journalism by a factor of nearly four.

The Virtual Newsroom Architecture

Data2Story doesn’t work by asking a language model to write a story about a dataset. It routes the task through a structured pipeline of seven specialized agents, each owning a discrete step in the journalism workflow:

• Detective — Searches the web for contextual background to frame the data
• Analyst — Executes code to compute figures and statistics from the raw CSV
• Editor — Determines which findings drive the narrative
• Designer — Selects appropriate media formats (maps, charts, audio elements)
• Programmer — Builds the HTML page
• Auditor — Reviews the layout for errors
• Inspector — Links every statement back to a verifiable source (Python code or external URL)

The Inspector is what drives the 93% verifiability figure. Rather than producing prose that cites sources loosely, each statement in the final output is either backed by runnable Python code that readers can execute themselves or linked to an external URL — making the evidence chain explicit and testable. According to The Decoder’s coverage, human-written data journalism from publications including The Economist and The Pudding achieved roughly 25% verifiability by the same standard.

What It Runs On

The system uses Claude Opus 4.7 running on Claude Code as its base model, with external models via OpenRouter for specialized media generation (images, video, and audio). The researchers demonstrated outputs on 2026 FIFA World Cup climate data, ArXiv submission trends, and time-use diary datasets, comparing results directly against comparable human-authored articles.

Why the Agent Separation Matters

The division of labor isn’t stylistic — it’s architectural. A single model handling the full data-to-article pipeline tends to drift: it loses track of which claims are data-supported versus synthesized, and verification becomes difficult to enforce post-hoc. By separating analysis (Analyst), narrative selection (Editor), and verification (Inspector) into distinct agents with bounded responsibilities, the system builds the evidence chain at construction time rather than trying to reconstruct it after the fact.

This is the same pattern that distinguishes production-ready multi-agent systems from single-model pipelines: task decomposition improves both reliability and auditability when the downstream output has verifiable correctness criteria.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.