Social computation is a paradigm in which a software application supports social interaction, but whose real purpose is the data trail that is left by that interaction, such as tags, recommendations, and so on. We explore a possible imperfection in the data that is generated by these games. Specifically, we investigate whether the data generated by previous participants influences the data that is generated by subsequent participants. We investigate this in the context of the ESP game. A feature of the ESP game is that words that have been generated by previous players are off-limits to subsequent players. The idea of this feature is to ensure that the system accumulates a variety of different words for each image. We consider the possibility that ironically, players are actually biased to suggest words that are related to the taboo words themselves. Based on anchoring and priming theories, we predict that the words players suggest will be related to the taboo words, and that this phenomenon will limit the variety of words that are collected for a given image. An empirical experiment confirms these predictions. This effect threatens to limit the potential value of socially generated information in many applications including recommender systems and ESP-like tagging systems, where later contributors are exposed to the inputs provided by earlier contributors.