Researchers built a chatbot for data journalists that is designed to be second-guessed, not obeyed.
The system, called DataWeave and detailed in a new arXiv paper, pairs conversational interaction with schema grounding, analytical planning, and query generation to help reporters explore large structured datasets. Rather than letting an LLM act as an autonomous answer engine, it keeps the model's reasoning and queries visible so a journalist can inspect, correct, and redirect them as a hypothesis shifts. The researchers tested it with professional journalists working on the U.S. Department of Education's IPEDS database, a sprawling, frequently-updated trove of postsecondary education data with its own thicket of coding conventions. The paper also serves as a deployment report, describing how real newsroom use reshaped the system's design.
The pitch addresses a well-documented failure mode of "ask in English, get SQL" tools: they misread domain-specific units, drift when a schema changes, and make silent assumptions a reporter would otherwise have to catch by hand. In data journalism, a wrong assumption about how a variable is coded does not just produce a bad chart - it can turn into a published false claim. Treating the model as a collaborator to argue with, rather than an oracle to trust, is a lower bar than most AI pitches, but it is the bar that actually matters in a newsroom.
It is a quieter promise than the usual AI-finds-your-story hype: not that the model surfaces the scoop, but that it is harder for the model to quietly get the scoop wrong.