Powering Trust: Tracking and Preserving Government Data References in the News Media

Abstract

Journalists frequently cite government data, yet poor citation practices and post-publication dataset modifications hinder independent verification (Camaj et al., 2025; Goes, 2025). This leaves data citations highly vulnerable to AI hallucination and misattribution.

Our paper analyzes 3,290 in-text citations from Canadian news media using a fine-tuned BERT model and RAG-powered AI agents to enrich data citations. To ensure integrity, we stored open data metadata and citations as a preservation layer on the Arweave blockchain, creating permanent, tamper-evident records of dataset versions via hashing. Our findings reveal that news media citations typically lack crucial metadata, such as versions and URLs, needed for verification. Crucially, we demonstrate how large language models (LLMs) can track citations across media, and how blockchain infrastructure provides the verifiability lacking in current government portals.

While focused on the Canadian context, this project offers an internationally transferable model for data accountability. Librarians are uniquely positioned to build this decentralized infrastructure. Serving as a critical verification layer for humans and LLMs, these workflows ensure digital collections remain trustworthy in a rapidly evolving technological landscape.

Keywords: data citation; open data verification; blockchain preservation; artificial intelligence; digital journalism