This paper investigates the feasibility of estimating current air quality conditions in cities without official air quality monitoring stations based on a statistical analysis of Twitter activity. For this purpose, a framework for collecting and geotagging air quality-related Twitter posts is developed and transfer learning is applied to enable estimations for unmonitored cities using data from monitored nearby cities. Experiments carried out on five cities in the UK and five cities in the US suggest that while Twitter-based estimates exhibit very high accuracy, they are outperformed on average by simple spatial interpolation. However, we find that a meta-model that combines estimates from spatial interpolation with Twitter-based ones increases accuracy in distantly located cities, highlighting the merits of Twitter-based air quality estimation and motivating further work on the topic.