Distributional semantic models (DSMs) are computational models that learn semantic relationships through word co-occurrence patterns, broadly aligning with human statistical learning mechanisms. Prior research has shown that DSMs capture not only general semantic structure but also human social biases. For example, Caliskan et al. (2017) demonstrated that pre-trained word embeddings encode associations that mirror implicit stereotypes measured by the Implicit Association Test (IAT). To better understands how DSMs acquire these biases, we examined the roles of two distinct sources of distributional information: first-order (direct co-occurrence) and second-order (indirect co-occurrence) statistics. Our analysis revealed that nearly all biases tested could be accounted for by first-order statistics alone, while about half were significant in second-order statistics. Every bias was present in at least one of these co-occurrence types, with nuanced variation in how different topics exhibited bias across first- and second-order associations. These findings suggest that implicit biases in DSMs can be attributed to simple co-occurrence patterns, predominantly direct associations. Moreover, they support theories positing that implicit biases reflect statistical regularities in the environment rather than personal attitudes. This work highlights how these biases are embedded in natural language and how a cognitive system capable of statistical learning could acquire implicit biases through the same mechanisms that shape human semantic memory.