During the unfolding of a crisis, it is crucial to determine its severity, yet access to reliable data is challenging. We investigate the relation between geolocated Tweet Intensity of initial COVID-19 related tweet at the beginning of the pandemic across Italian, Spanish and USA regions and mortality in the region a month later. We find significant proportionality between early social media reaction and the cumulative number of COVID-19 deaths almost a month later. Our findings suggest that "the crowds" perceived the risk correctly. This is one of the few examples where the "wisdom of crowds" can be quantified and applied in practice. This can be used to create real-time alert systems that could be of help for crisis-management and intervention, especially in developing countries. Such systems could contribute to inform fast-response policy making at early stages of a crisis.