AI's Influence on Socially Constructed Kinds

Abstract

This paper examines how large language models (LLMs) using natural language understanding (NLU) distort the framework of looping effects and interactive kinds as characterized by Ian Hacking. Drawing on Hacking's classificatory practices, Laimann's capricious kinds, and Tekin's cyclical exchange framework, the paper demonstrates how contemporary AI systems produce inconsistent results when addressing socially constructed classifications — increasing stigmatization surrounding social issues and amplifying the capricious nature of interactive kinds. Through a comparative case study of a user researching a Schizophrenia diagnosis across Google search and tiered ChatGPT models (Legacy GPT-3.5, Default GPT-3.5, and GPT-4), the paper shows how AI's opacity, lack of source transparency, and paid-tier disparities introduce biased conceptualizations into the feedback loop between classification and classified people. Note: This paper presents a case study and empirical data alongside its companion piece, "Robots in Disguise" (2024), which extends the analysis to normative ethics and prescriptive codes of conduct for generative AI organizations. The two papers are intended to be complementary: this paper provides the descriptive and analytical framework, while the companion paper presents the normative and prescriptive argument.

Ähnliche Arbeiten