Translating control: Benchmarking language AI and sociotechnical governance

In this article, I examine the rise of AI-assisted translation and interpretation in healthcare as a sociotechnical governance problem rather than a purely technical one. Drawing on ethnographic research in Oregon, U.S., I show how automated language tools are used in clinical settings despite regulatory frameworks that prioritize certified human interpreters. Patients and clinicians often turn to tools such as Google Translate and other AI-assisted platforms as pragmatic responses to institutional delays, staffing shortages, and workflow pressures, redistributing communicative trust across devices. I argue that evaluation frameworks for language AI, rooted in benchmark metrics like BLEU or COMET in machine translation, are epistemologically misaligned with the relational, contextual, and ethical nature of healthcare interpretation. I conclude by advocating a sociotechnical governance approach that treats indeterminacy as a condition to be managed through layered oversight, risk-based workflows, and contextual accountability, rather than eliminated through technical optimization.Implications Translation and interpretation function not as acts of equivalence, but as situated mediation involving judgment, affective labor, and risk management, none of which can be adequately captured by standardized metrics in Language AI. Rather than framing automation as replacement or failure, I highlight a crisis of governance: responsibility for communication quality is increasingly displaced from institutions onto frontline workers and patients.Social Media Statement When a patient turns to Google Translate rather than human interpreters, who is actually responsible for the medical failure when AI mediates healthcare communication?