“Argh! The World Doesn't Fit the Model!”

This article examines how data annotation in news organizations serves as a site of friction and world-making, shaping how AI systems categorize and interpret journalism. Through an ethnographic study in a Danish news organization, we identify three interrelated sites of friction: taxonomic constructions, organizational infrastructures, and aspirational realities. Rather than a neutral task, annotation is a contested process where journalistic epistemologies and machine learning logics collide, producing small acts of world-making that influence how journalistic knowledge is computationally instantiated. By centering friction as an analytical lens, we extend the infrastructural turn in media theory, showing how annotation practices – caught in between the journalistic and the computational domain - themselves function as infrastructural sites of negotiation and power. We thus show how AI is not seamlessly integrated into news organizations but shaped through ongoing struggles over classification, professional authority, and epistemic legitimacy, highlighting the infrastructural politics that underpin the computational turn in journalism.