The monolingual suspicion: AI detectors, translingual authorship and the assessment of multilingual academic writing

Abstract

This theoretical essay examines the use of AI-generated text detectors in assessing the academic writing of multilingual writers and asks which conception of language and authorship becomes institutional rule when a probabilistic score is treated as evidence. Drawing on Critical Applied Linguistics, translanguaging theory, Ubuntu translanguaging and the raciolinguistic perspective, the analysis combines a critical reading of empirical studies on detectors published between 2023 and 2026, including those that found no bias, an analysis of Brazilian legislation on assessing the writing of deaf people, and four theoretical cases involving deaf, Indigenous and refugee students and researchers who write in English. The essay holds that detector behaviour depends on language and calibration and argues that the institutional use of scores, combined with deficit expectations, produces a monolingual suspicion. In this configuration, a text that retains the marks of the writer's repertoire tends to be read as deficient, a text close to the norm tends to be read as inauthentic, and the burden of proof falls on the writer. The essay reads Decree No. 5,626/2005 and the Brazilian Inclusion Law as a precedent for non-monolingual assessment and proposes a relational translingual authorship, with a gradient of mediations and five assessment criteria.