Data2Story: CSV-to-article pipeline with seven AI agents

Data2Story, built by researchers from Oxford and Stanford, turns a CSV file into a full interactive online article automatically using a Claude Code skill and seven specialized agents. The pipeline produces narrative text, charts, multimedia and an "Inspector" panel that links 93 percent of visible statements to runnable code or external sources.

What is Data2Story and how does it work?

Data2Story is a Claude Code skill called "Data Journalist Agent" that loads a predefined task set and orchestrates seven specialist agents to convert a dataset into a verifiable, multimodal web article. The system runs on Claude Opus 4.7 via Claude Code, and for images, video, and audio it pulls OpenRouter models such as gpt-5.4-image-2, seedance-2.0, and lyria-3-pro-preview.

The pipeline begins with a CSV file and a "virtual newsroom" of agent roles. The Detective runs web searches for context and links external sources. The Analyst runs code to compute figures from the data. The Editor selects which findings form the narrative. The Designer picks presentation media. The Programmer builds the HTML page. The Auditor checks layout and errors. The Inspector creates index cards tying each sentence, chart or interactive element to either a script plus data file or an external URL.

How did it perform in tests against human articles?

In an evaluation using 18 public datasets paired with matching human-written originals, 53 recruited readers rated agent versus human pieces across five categories and preferred the agent version overall. Data2Story won all five categories; transparency led the margins at +1.49 on a seven-point scale. Overall preference was 74 percent for the agent article, 25 percent for the human original, and 2 percent a draw.

The team used datasets and human comparisons drawn from The Economist briefings, The Pudding long reads, and TidyTuesday community datasets. By source, the agent outperformed humans clearly on data-heavy Economist briefings and TidyTuesday pieces, while against Pudding reports it was a statistical tie. Across the 18 article pairs the agent reproduced about half of the human findings, while only 35 percent of the agent's statements appeared in the human texts.

The authors demonstrate Data2Story on a number of datasets, including a climate-focused article built from the 2026 FIFA World Cup schedule and host cities. From that schedule the system notes "about four in ten matches" are slated for locations the players' union FIFPRO classifies as extremely high heat risk, and that humidity, rather than air temperature, is the main driver; the researchers emphasize those are typical climate conditions, not a forecast for the tournament.

Why it matters

Data2Story replaces much of the mechanical work of data journalism with an auditable, runnable pipeline: the Inspector links claims to either the exact line of code plus the data file or to an external URL, making 93 percent of visible statements checkable compared with a 25 percent baseline for human-written pieces. That gap highlights both a reproducibility deficit in current journalism practice and a concrete method for reducing "attribution hallucination" by producing runnable calculations rather than opaque assertions.

The system is not positioned as a replacement for reporting. The researchers flag three areas where humans retain an edge: editorial perspective and explanatory reporting, bespoke creative design, and dense single graphics that pack layered annotations. The team frames the tool as a collaborator that can surface niche or undercovered datasets that newsrooms lack time to investigate.

What to watch

A human-in-the-loop version is left for future work; the current implementation runs on full autopilot. The project is live at data2story.github.io, and the code is available on GitHub. Watch for a published human-in-the-loop release or follow-up evaluations that test whether editorial oversight closes the gap on perspective and creative presentation.