Talking with Machines: Charting the Space of GenAI Practices

• In Event: 41010 - Regular Session on Culture and Inequality

Abstract

Generative AI tools such as ChatGPT are increasingly woven into everyday routines, assisting with everything from code debugging to recipe brainstorming and medical queries. Far from mere convenience, these interactions shed light on underlying cultural and social dynamics. Drawing on Bourdieu’s (1979, 1980) insights into habitus and Swidler’s (1986) concept of cultural repertoires, we posit that user–AI dialogues reflect the ways people navigate social position. For instance, occupational norms may color perceptions of which requests are legitimate (e.g., coding fixes) or risky (e.g., medical advice), while national contexts can influence linguistic strategies—polite hedges versus direct commands.

To systematically explore this terrain, we analyze one million anonymized ChatGPT conversation logs from the WildChat Database (Zhao et al. 2024), focusing on English-speaking users across Canada, South Africa, India, the United States, the United Kingdom, Australia, and New Zealand. First, we parse each session into “task requests” using a fine-tuned BERT model (Devlin et al. 2019), allowing multi-label classification to capture overlapping uses (e.g., coding plus recipe queries). We then employ cluster analysis and factor analysis (Le Roux and Rouanet 2004; Sapiro 2023) to map the co-occurrence of tasks, distinguishing “mainstream” from atypical or “marginal” uses. Integrating these findings with user-reported or inferred social data—occupation, location—we examine whether higher-status individuals, for example, treat AI with more or less deference, and whether certain task combinations reflect or defy established social boundaries.

Lastly, we conduct a sociolinguistic analysis of prompt styles, probing whether people persist in using politeness markers or adopt blunt, command-like phrasing when engaging an ostensibly non-human agent. By situating AI usage within a broader field of cultural practice, this study highlights how even mundane interactions with generative models can illuminate deeper structures of taste, power, and social distinction in the digital era.

Authors

• Leo Paul Henry, University of Toronto

• Alexander Wilson, University of Toronto