People often have thoughts, attitudes and biases that are not themselves consciously aware of or that they would rather not share with others. To assess such attitudes, researchers use paradigms like the Implicit Association Test (IAT) that do not rely on explicit responding to determine the level of bias a person holds towards a particular target concept (e.g., race, gender, age). Responses in the IAT are assumed to reflect deeply held beliefs and attitudes, and not shallow, superficial associations. However, as linguistic distributional information has been shown to serve as a viable heuristic in many cognitive tasks, we investigated whether it could be used to predict the level of bias established by the IAT. We used a large corpus of language (Web 1T) and data from 16 IAT studies (N = 1825) to examine whether the degree of linguistic co-occurrence for target concepts and attributes reflected the size of bias observed in human behavioural data. We found that the effect size of the linguistic biases corresponded strongly with the effect sizes from the behavioural data. We suggest that language reflects prevalent cultural attitudes which are captured by tasks such as the IAT, suggesting that the IAT may reflect shallow, linguistic associations rather than deeper conceptual processing.