The most tedious, time-eating work in journalism isn’t writing — it’s the invisible machinery of verification. Fact-checking, sourcing, recalculating statistics, tracing a single claim back to its origin in a spreadsheet. A team of researchers from Oxford and Stanford just automated that entire chain with a system that takes a raw CSV and spits out a fully interactive, verifiable news article. And readers prefer what the agents produce.
Data2Story isn’t a single model generating text. It’s a seven-agent virtual newsroom that mirrors a real editorial workflow. There’s a Detective that runs web searches for context, an Analyst that writes and executes code instead of hallucinating numbers, an Editor that selects the narrative arc, a Designer that picks the right visual medium, a Programmer that builds the HTML page, an Auditor that checks layout, and an Inspector that ties every single sentence, chart, and interactive element back to its source. The base model is Claude Opus 4.7 running on Claude Code, with OpenRouter models pulled in for images, video, and audio.
The real breakthrough isn’t the output quality, though that surprised the researchers too. It’s the Inspector panel that achieves 93% source-link coverage. Every claim gets an evidence card: either the exact line of code that computed the figure, or the external URL that backs it. That’s not correctness — the researchers are careful to say it’s verifiability. If you doubt a number, you can run the script yourself. For human-written articles, the baseline is 25%, largely because journalists don’t publish their analysis code. The gap is an indictment of current practice, and Data2Story closes it by design.
This matters because attribution hallucination is the defining trust problem of AI-generated content. Models often give the right answer but cite the wrong source, a failure mode that plagues retrieval-augmented systems. Data2Story sidesteps that by having the Analyst compute figures from the data itself with runnable code, and by hard-linking every output to its provenance. It’s the first end-to-end system I’ve seen that treats verifiability as a first-class architectural constraint, not a post-hoc filter.
In a blind study with 53 recruited readers and 18 dataset-article pairs from The Economist, The Pudding, and TidyTuesday, the agent articles won in all five rated categories: visual design, narrative rhythm, data transparency, verifiability, and insight gained. The biggest lead was in transparency, at +1.49 on a seven-point scale. Overall, 74% of readers preferred the agent article. Against the lavishly handcrafted long-reads of The Pudding, it was a statistical tie, which is staggering when you consider The Pudding spends weeks on a single piece.
Yet the system’s weaknesses reveal where human journalists still dominate. The agent can tell you what is in the data, but not why. A human report on repair cafes explains that manufacturers deliberately block access to diagnostic tools — an investigative theory grounded in reporting, not spreadsheets. The agent charts repair rates by product type and stops there. On creative design, it can’t match the bespoke interfaces The Pudding builds, like turning a comedy transcript into an interactive timeline of laugh lengths. And when the data demands a single, densely annotated graphic, the agent scatters insights across multiple charts and loses the point.
For builders, the architecture is the real instruction manual. The seven-role decomposition is not arbitrary; it mirrors the cognitive load that attribution demands. The Detective researches, the Analyst computes, the Editor decides what matters, the Inspector verifies. Each agent has a narrow, testable responsibility, and the handoffs are explicit. That’s the pattern to steal: when you need an output that must be backed by evidence, break the pipeline into roles that can prove their work independently. The Inspector isn’t an afterthought — it’s the spine.
There’s also a fascinating asymmetry in coverage. Data2Story captures roughly half of the human article’s statements, but human articles capture only 35% of the agent’s. The agent adds plenty of its own angles, especially in data-heavy briefings, yet misses the editorial core that requires external knowledge. This suggests a hybrid future where the agent drafts the data-rich skeleton and the human injects interpretation, narrative, and context. Full autopilot is a parlor trick; human-in-the-loop is where the real newsroom of tomorrow lives.
I’m less interested in whether Data2Story replaces journalists than in what it replaces. It won’t do investigations or cultural criticism, but it will swallow the commodity news grind — the quarterly earnings recaps, the sports stat rundowns, the environmental data dumps that newsrooms can’t cover because there aren’t enough bodies. That’s a net gain. A verifiable, machine-generated article on FIFA World Cup heat risk is better than no article at all. And when the Inspector panel gives readers a button to see exactly how every number was produced, we might even end up with journalism that’s more trustworthy than the human kind.
