Computational methods, and in particular large embedding models, represented a breakthrough in machine learning and artificial intelligence by using large amounts of training data to capture nuanced relationships and meaning in text and images. At the same time, they surfaced a pressing dilemma: these models encode societal biases and stereotypes inherent in that training data. While machine learning practitioners are addressing the ethical concerns of producing biased models, social scientists and humanists are celebrating these models’ faithful encoding of social biases as a means to study, analyze, and, for some, objectify societal constructs. The adoption of these sometimes opaque models into the social sciences and humanities, however, has raised old and new epistemological challenges, in particular by challenging traditional methodological binaries between the subjective and objective, the universal and partial. While some scholars have argued that these models fully realize the ideal of subjectless objectivity, others claim they realize Kantian ideals around universal knowledge via intersubjective validity. This chapter instead argues that these models operationalize Donna Haraway's scientific ideal of objectivity through situated knowledges and the partial perspective. If computational models encode social constructs, they necessarily encode the multiple perspectives present in social data. Harnessing computational methods to achieve objectivity via partial perspectives, however, requires a focus on data provenance, to ensure the validity and fidelity of computational output. I end with a call to revive data provenance in the newest iterations of computational modeling.
