The impacts of environmental change on human outcomes often depend on local exposures and behavioral responses that are challenging to observe with traditional administrative or sensor data. We show how data from private pollution sensors, cell phones, social media posts, and internet search activity yield new insights on exposures and behavioral responses during large wildfire smoke events across the US, a rapidly-growing environmental stressor. Health-protective behavior, mobility, and sentiment all respond to increasing ambient wildfire smoke concentrations, but responses differ by income. Indoor pollution monitors provide starkly different estimates of likely personal exposure during smoke events than would be inferred from traditional ambient outdoor sensors, with similar outdoor pollution levels generating >20x differences in average indoor PM2.5 concentrations. Our results suggest that the current policy reliance on self protection to mitigate health risks in the face of rising smoke exposure will result in modest and unequal benefits.