‘Glazing Models’: sycophancy and the dynamics of synthetic feedback

Jones, R. ORCID: https://orcid.org/0000-0002-9426-727X (2025) ‘Glazing Models’: sycophancy and the dynamics of synthetic feedback. In: Vasquez, C. and Jaworska, S. ORCID: https://orcid.org/0000-0001-7465-2245 (eds.) Feedback Society. Routledge, Adington, Oxon. (In Press)

Abstract/Summary

This chapter uses sociolinguistic analysis to understand how humans and AI co-construct relationships through feedback interactions. Drawing on DuBois's stance triangle framework, it analyses Reddit discussions of the April 2025 scandal surrounding ChatGPT 4o's sudden increase in sycophantic behaviour. The analysis shows how ‘synthetic feedback’ from large language models involves complex negotiations of stance that position both users and AI as particular kinds of social actors. In their discussion of AI generated feedback, users developed detailed folk taxonomies of sycophantic linguistic features and associated these with specific characterological figures, demonstrating how AI stancetaking behaviours become enregistered through collaborative online discourse. While users criticized ChatGPT's excessive agreeableness, their proposed solutions—emphasizing ‘cold’, emotionless feedback—also reflect problematic ideological assumptions about feedback that ignore its inherently relational dimensions. The study argues that understanding AI sycophancy requires examining not just model outputs but the ongoing metapragmatic negotiations between humans and machines which shape their broader sociotechnical imaginaries about AI capabilities and appropriate human-AI relationships. Keywords: AI, enregisterment, metapragmatics, stance, sycophancy, synthetic feedback

Item Type	Book or Report Section
URI	https://centaur.reading.ac.uk/id/eprint/124257
Refereed	Yes
Divisions	Arts, Humanities and Social Science > School of Literature and Languages > English Language and Applied Linguistics
Publisher	Routledge
Download/View statistics	View download statistics for this item