Ambient AI Scribes as Emerging Infrastructure in the Learning Health System

ABSTRACT

Ambient artificial intelligence (AI) scribes are systems that automatically generate clinical documentation from clinician‐patient conversations and are being deployed at accelerating pace across US health systems. Early evaluations report reduced documentation burden, improved clinician well‐being, and perceived efficiency gains, reinforcing a narrative of inevitability. Yet this frontline framing understates a more consequential issue: ambient scribes outsource the “first mile” of clinical documentation, thereby reshaping the production of clinical data and the learning health systems (LHSs) that depend on documentation as foundational infrastructure. This paper argues that ambient AI scribes should be understood not merely as workflow tools, but as emerging infrastructure that will materially shape the capacity and capabilities of LHSs. Drawing on infrastructure studies and LHS frameworks, we conceptualize clinical documentation as the epistemic substrate through which encounters are translated into analyzable data that power quality measurement, predictive modeling, clinical decision support, and institutional learning. When this translation is algorithmically mediated by proprietary systems, design choices, training data, and integration pathways can introduce systematic documentation errors that propagate downstream, often invisibly, through analytic pipelines. Synthesizing emerging evidence, we highlight risks including hallucinated clinical details, omission of safety‐critical information, and differential performance across patient populations with diverse accents or speech patterns. These risks mirror classic infrastructural properties described by Star: embeddedness, dependence on the installed base, wide propagation, and visibility primarily upon breakdown. From this perspective, ambient scribes may quietly reshape documentation norms, data quality, and learning trajectories well before downstream effects are routinely assessed. We conclude by outlining a governance agenda grounded in LHS principles: documentation‐quality metrics, drift monitoring, equity‐focused evaluation, transparency, and multi‐stakeholder stewardship. Without such oversight, ambient AI scribes risk stabilizing an infrastructural layer that delivers short‐term relief while eroding the long‐term integrity, equity, and trustworthiness of learning health systems.