Object recognition is a central function of the visual system. As a first step, the features of an object are registered; these independently encoded features are then bound together to form a single representation. Here we investigate the locus of this ‘feature integration’ by examining crowding, a striking breakdown of this process. Crowding, an inability to identify a peripheral target surrounded by flankers, results from ‘excessive integration’ of target and flanker features. We presented a standard crowding display with a target C flanked by four flanker C’s in the periphery. We then masked only the flankers (but not the target) with one of three kinds of masks – noise, metacontrast, and object substitution – each of which interferes at progressively higher levels of visual processing. With noise and metacontrast masks (low-level masking), the crowded target was recovered, whereas with object substitution masks (high-level masking), it was not. This places a clear upper bound on the locus of interference in crowding suggesting that crowding is not a low-level phenomenon. We conclude that feature integration, which underlies crowding, occurs prior to the locus of object substitution masking. Further, our results indicate that the integrity of the flankers, but not their identification, is crucial for crowding to occur.