The harsh lockdown measures that marked the response to the COVID-19 outbreak in theItalian region of Lombardy provides a unique natural experiment for assessing the sensitivity oflocal air pollution to emissions. However, evaluating the pollution benefits of the lockdown iscomplicated by confounding factors such as variations in weather. We use a machine learningalgorithm that does not require identifying comparable but unaffected regions while addressingthe effect of weather. We show that the lockdown, albeit virtually halting most human ac-tivities, reduced background concentrations of PM2.5by 3.84μg/m3(16%) and NO2by 10.85μg/m3(33%). Improved air quality has saved at least 11% of the years of life lost and 19%of the premature deaths attributable to COVID-19 in the region. Although air pollution hassignificantly decreased, it has often remained above safety thresholds. The analysis highlightsthe diversity of air pollution sources and the need for an expansive policy response.