AI Journalism Study: Multi-Agent Systems Hit 84.7% Accuracy but Take Twice as Long

A new study from researchers Obada Kraishan, Kulsawasd Jitkajornwanich, and Kerk Kee presents the first systematic comparison of four distinct AI agent architectures for computational journalism. Using gatekeeping theory as a lens, they tested monolithic (Claude), chain-based (LangChain), multi-agent collaborative (CrewAI), and autonomous iterative (AutoGPT) systems across 200 controlled experiments spanning 50 journalism tasks. All architectures used the same underlying language model and identical tools, isolating the effects of architectural design alone.

The results reveal significant performance tradeoffs. Multi-agent collaboration achieved the highest accuracy at 84.7%, but required roughly twice the time of other architectures. The monolithic architecture exhibited a 71.7% source rejection rate—a quantitative parallel to classic human gatekeeping in newsrooms. Chain-based designs excelled in speed and structured attribution, monolithic in versatility, and iterative in auditability and methodological documentation. Architecture explained 82% of variance in processing behavior, positioning it as a new structural level of gatekeeping. The study provides evidence-based guidance for newsrooms: chain for speed, multi-agent for accuracy, monolithic for general tasks, and iterative for transparency.