AI still has a vocabulary problem: toward terminological accountability in artificial intelligence

Abstract

The artificial intelligence (AI) industry’s vocabulary–including “hallucination”, “frontier AI”, “the cloud”, “responsible AI”, “benefiting humanity”--has been examined through individual terms in prior research. This work examines these terms as an interlocking vocabulary and investigates how they help insulate the organizations with greatest control over AI from accountability for what that control costs others, which we call “accountability insulation”. We synthesize thirty terms, grouping them by the questions of responsibility each term deflects: who is responsible for the output, who bears the costs, how speculative futures obscure present harms, which progressive values are co-opted, and what fundamental assumptions go unquestioned. We discuss these terms in relation to ideological commitments discussed in prior work, including those associated with the TESCREAL bundle and techno-utopian ideologies that help explain why these terms return in new forms, and to research on the labor exploitation, environmental impacts, and social harm that accumulate beneath the promises these words make. Behind the language of seamless intelligence are workers who classify traumatic content for poorly paid wages, communities whose water is consumed by server farms, and individuals whose lives are determined by probabilistic decision tools sometimes named, reassuringly, “solutions”. This paper addresses AI researchers, reviewers, policymakers, and the public, each of whom encounters and relies upon AI vocabulary. We conclude with a call for terminological accountability, organized around the Resistance, Refusal, Reclamation, and Reimagining of harmful vocabulary, offering guidance on navigating these terms where they risk contributing to diminished recognition of those already harmed by AI technologies.