Contemporary artificial intelligence (AI) technologies are often presumed to be capable of revealing unmediated truths about the world, including the truths language might hold, echoing the long‐standing assertion that language's primary function is to directly translate reality. These interrelated representationalist ideologies unite in vocal biomarker AI, a subarea of machine learning research and development dedicated to optimizing how the voices of people experiencing mental distress are interpreted. Although the premise behind vocal biomarker AI – to listen to the brain through the voice – assigns transcendent semiotic capacities to both AI and language, another story unfolds in practice. Drawing from fieldwork with vocal biomarker AI researchers, I show that researchers take the social, pragmatic, and political dimensions of language to be foundational to its computational modelling. This is because to gather the voice data necessary to their research, researchers must recruit human research subjects and make them speak, contending with the gaps between the subjects’ and their own expectations regarding the meanings and values that language can embody. By underscoring the centrality of disjuncture, maintenance, and partiality – i.e., transduction – in vocal biomarker AI, I counter hegemonic claims about the people behind AI systems and the systems’ abilities to do things with words.
Beth M. Semel (Mon,) studied this question.