27.5. REPAIR Dialogues with Wanheng Hu

Postdoctoral Scholar Wanheng Hu (Stanford University) will deliver an online presentation titled When Accuracy Falters: Repair Practices and the Code–Data Equilibrium in Medical Imaging AI on May 27, from 9:00 to 10:00 (EET).

Attend the session via Zoom: [link will be added closer to the event]

REPAIR Dialogues is a venue for scholars and collaborators of the REPAIR project to present their ongoing work and findings in a less formal manner. Each event includes an introduction by a researcher or expert, followed by a joint discussion. The events are held online (via Zoom) in English.

When Accuracy Falters: Repair Practices and the Code–Data Equilibrium in Medical Imaging AI

Abstract: Machine learning (ML) models for medical image analysis are commonly evaluated through a set of quantitative performance metrics centered on accuracy, often treated as objective indicators of the model’s inherent technical properties. Drawing on ethnographic research in two Chinese medical AI companies, this talk examines moments when model performance falters and the subsequent attempts to “repair” the system during the training stage. I show how ML engineers and developers respond to performance breakdowns by re-aligning datasets, annotation protocols, model architectures, and evaluation procedures—practices that illuminate the sociotechnical terrain on which accuracy depends. I conceptualize these efforts as the maintenance of a code–data equilibrium, a dynamic balance through which a model comes to appear accurate, stable, and credible. Moments when this equilibrium breaks down, particularly in internal testing, reveal how engineers navigate the regress between code and data, weighing the credibility of annotators, the cost of checking data, available modeling techniques, and project time constraints as they determine where repair should occur. By reframing repair as the continual re-stabilization of the sociomaterial conditions that make algorithmic performance possible, the talk offers a grounded perspective on the vulnerabilities, maintenance labor, and forms of expertise embedded in contemporary AI systems.

Wanheng Hu is a scholar of Science and Technology Studies (STS) whose research examines the epistemic, ethical, and regulatory dimensions of artificial intelligence. His current book project, Reassembling Expertise: Credible Knowledge and Machine Learning in Medical Imaging, draws on multi-sited ethnography in China’s medical AI industry to analyze how medical expertise is translated into AI systems and how credibility is negotiated across industrial, clinical, and regulatory settings. 

Wanheng is currently an Embedded Ethics Fellow at Stanford University’s McCoy Family Center for Ethics in Society and the Institute for Human-Centered AI (HAI), and will join Peking University as Assistant Professor of STS in summer 2026.