Trained Judgements Artificial Intelligence, Epistemic Tensions and the Production of Scientific Objectivity

In this paper, we investigate uses of AI (Artificial intelligence) in two distinct fields: radiology and prehistoric archaeology. We examine the normative tensions between the scripts encapsulated within the technology and pre-existing professional and epistemic cultures, as well as the situations in which mechanical objectivity fits with local norms. Through ethnographic observation and interviews in French field sites, we show how in radiology a specific definition of “normal” bodies, embedded within detection tools, conflicts with medical practice, and the way in which non-consensual knowledge in archaeology can challenge the prediction of soil occupation in a prehistoric site. We also highlight the conditions under which AI tools can adhere to certain epistemic norms and become part of professional practices in radiology and prehistoric archaeology. While in radiology AI is judged by its ability to close uncertainties without imposing binary categories, in prehistoric archaeology, its epistemic validity depends on mobilizing exogenous scientific data to increase researchers’ reflexivity about their practices and knowledge, suggesting new clues and explanatory paths. This article demonstrates the effectiveness of AI technologies is shaped by local constraints, and why their objectivity is not a given property but an emergent feature arising from specific contexts of use.