From Black Box to Accountable Sensor: A Reliability-Aware Generative AI Framework for Public-Facing Wearable Health Devices
Main article
Abstract
Consumer wearables increasingly combine biosignal sensing, generative signal enhancement, and automated health interpretation. This convergence creates a new socio-technical problem: a device that appears to be a simple sensor may actually contain a generative model that edits, completes, or denoises the signal before a downstream classifier or large language interface turns it into advice. When the generated signal is reliable, such adaptation may improve public access to early health warnings. When it is unreliable, the same mechanism may transform a noisy input into a confident but misleading output. This article develops a Reliability-Aware Generative AI (RA-GAI) framework for public-facing wearable health devices. The framework combines signal quality assessment, generative adaptation, decision-theoretic uncertainty, action gating, and an accountability layer that records why a device produced, withheld, or escalated a health interpretation. Using a photoplethysmography-oriented wearable scenario, the article presents a transparent reliability simulation that compares raw inference, ungated generative enhancement, uncertainty-gated enhancement, and the full accountability framework across noise, domain shift, and deployment-risk conditions. The results indicate that generative adaptation can improve the apparent performance of health inference, but its public safety value depends on whether uncertainty is translated into action-specific controls. The proposed framework improves accepted-window accuracy, reduces unsafe automatic alerts, and clarifies when a device should return a user-facing caution rather than a diagnosis-like claim. The contribution is twofold: technically, the article converts uncertainty from a model-internal statistic into an operational reliability signal; socially, it defines accountable sensing as a governance requirement for wearable technologies that intervene directly in everyday health decisions.
