The social origins of identity and bias: Language models, accumulated meaning and positioning

This paper develops a social psychological interpretation of distributional language models, algorithmically learned representations of natural language. Within these models, words are positioned according to patterns of usage, and the resulting embeddings can be interpreted as maps of cultural associations available for learning. This cultural-learning account explains how patterned meanings become available to individual cognition but leaves open how those meanings are socially produced. Drawing on social identity theory, discursive psychology and a stigmergic-dialogical model of communication, the paper argues that identity and bias emerge through social interaction and the accumulation of linguistic traces. As people take positions, respond to others and circulate formulations, they produce patterns of meaning that locate actors relationally and make alignments and differences recognizable. Bias emerges when this accumulation is uneven, making associations, evaluations and distinctions differentially available across communities. From this perspective, embeddings map patterns accumulated through social activity. Their geometry captures relations through which identities become recognizable and groups distinguishable. The paper extends this approach with speaker landscapes, which locate speakers by their language use and show how groups become differentiated. The result is a framework for studying how meaning, identity, bias and group differentiation are socially produced at scale.