MBZUAI's QRAFT Puts AI Fact-Checkers to the Test, With Human Editors as the Judge

Three AI agents, each assigned a distinct newsroom role, now form the backbone of a system built to answer one of journalism's hardest questions: can a machine write a credible fact-check?

Researchers at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) have introduced QRAFT, an agentic framework that attempts to replicate how professional fact-checkers actually work rather than simply summarizing claims and verdicts. The system splits the task among a planner agent, a writer agent, and an editor agent. These three components don't just pass work down an assembly line; they interact through conversational question-and-answer exchanges, mirroring the back-and-forth that occurs in real editorial workflows when a draft gets challenged, revised, and refined before publication.

That design choice stems directly from fieldwork. Before building QRAFT, the MBZUAI team interviewed working fact-checkers at established fact-checking organizations to map out how these articles are actually constructed, from initial claim identification through sourcing and final editing. Dhruv Sahnan, a doctoral student in Natural Language Processing at MBZUAI, is among the researchers behind the project, alongside a notably cross-disciplinary group of co-authors: David Corney, Irene Larraz, Giovanni Zagni, Ruben Miguez, Zhuohan Xie, Iryna Gurevych, Elizabeth Churchill, Tanmoy Chakraborty, and Preslav Nakov. The mix of academic NLP researchers and figures tied to fact-checking practice reflects the project's core premise: that automating this kind of writing requires input from the people who already do it professionally.

Testing Against Human Judgment

The team didn't rely solely on automatic scoring to validate QRAFT. Alongside standard metrics used to evaluate machine-generated text, the researchers brought in professional fact-checkers to manually rate the articles QRAFT produced. That dual-track evaluation, machine metrics plus human review, gives the results more weight than a purely algorithmic benchmark would. On the automatic measures, QRAFT outperformed comparison approaches, suggesting the multi-agent, conversational structure produces text that scores better on established quality indicators for generated content.

The study was presented at the 64th Annual Meeting of the Association for Computational Linguistics (ACL) in San Diego, one of the most prominent venues in computational linguistics research, lending the work a degree of peer scrutiny beyond an internal lab release.

Why This Matters Beyond the Lab

Automated fact-checking tools have circulated for years, but most focus narrowly on classifying a claim as true or false. QRAFT's ambition is different: it targets the harder task of producing the actual explanatory article, the kind of nuanced, sourced writing that fact-checking organizations publish daily to walk readers through evidence and reasoning. If agentic systems like this can reliably approximate that process, newsrooms facing growing volumes of viral misinformation could gain a meaningful assist, provided human editors remain in the loop to catch what automated judgment still misses.