Performative Risk-Sensitive Routing for Personalized Health Navigation

Generative health assistants increasingly support personalized planning, while pedestrian routing remains largely optimized for time and distance. A natural extension is to recommend routes that reduce exposure to air pollution, allergens, or crowded poorly ventilated spaces. We argue that independently optimizing such routes is incomplete in dense cities: recommendations can change pedestrian flow and therefore change part of the exposure field being optimized. Health-aware navigation is thus \emph{performative}. We propose a modular methodology combining mostly exogenous epidemic and atmospheric forecasts with a policy-dependent crowd model, a contextual exposure model, private user preferences, risk-sensitive route evaluation, and population-level coordination. A generalized two-route model shows that individually rational health avoidance can over-concentrate users on a nominally safer route; for crowd cost $g(x)$, the ignored marginal social cost is $xg'(x)$. Generative AI serves as a bounded preference parser, tool orchestrator, and explanation layer rather than as an unconstrained epidemiological predictor or route generator. COVID-era mobility motivates endogenous respiratory-exposure risk, while allergic-rhinitis and pollution scenarios illustrate broader personalized use cases.