In 2020, Japan introduced a comprehensive heat-health warning system where daily alerts were issued by region when forecasted wet bulb globe temperature (WBGT) exceeded a threshold (33 °C). Utilizing plausibly exogenous region-day variation in the difference between actual and forecasted WBGT (i.e. forecasting errors), we find that the alerts led to a large and precisely estimated increase in heat stroke counts. Paired with data from Google Trends, Google Mobility Reports, and the population of ambulance records, we identify potential mechanisms, including increased reporting of heat stroke cases and "adverse" behavioral responses (e.g. people spending more time outdoors) when alerts were issued, while ruling out potential substitution in health diagnoses away from other sudden illnesses.