Abstract
Abstract
Large language models produce speech that users recognise as human. That recognition is prepared before any output, in the documents that instruct the workers who train and evaluate these systems. Research on language models has concentrated on their outputs and on their technical architectures, which leaves the discursive infrastructure preceding both largely unexamined. This article analyses four training guidelines addressed to annotators and quality evaluators of commercial language models, obtained through the fieldwork of one of the authors, who worked in both roles. Through critical discourse analysis, the documents are read for the rules of correct speech they establish and for the way those rules translate human language into classification matrices. The analysis reconstructs two dimensions of the normative regime that the guidelines organise. The first, arranged around the triad of helpfulness, honesty and harmlessness, fixes the semantic and moral conditions of correct speech, defining what can be said, from which epistemic position, and within which ethical limits, with harmlessness receiving the most detailed operationalisation. The second dimension extends that regime to tone and interaction, regulating how much clarity and courtesy a response must carry, along with the sonorous texture of the generated voice. Taken together, the guidelines assemble a model of human communication by subtraction, since they drain the ideal interlocutor of everything that situates real speakers in a body, a history and a position. The humanity that these systems reproduce is therefore fabricated within a hierarchised labour process where the decision about which behaviours count as acceptable precedes any technical property of the model.