Back to search

University of Waterloo

How social context shapes value-related content in Large Language Model outputs

Abstract

dc:description.abstract

Large language models (LLMs) are increasingly used in settings where their output must align with specific values, goals, or preferences. Whether the text produced by these models reliably reflects the values they are instructed to follow, and whether another model can identify those values from the generated text, are open questions in value alignment research. Most prior work evaluates value alignment by comparing model outputs to human-labelled data collected in settings where the social context is not systematically varied, leaving open the question of how the social context of an interaction affects whether a model's output reflects the intended value and whether that value can be recognized from the text. In this thesis, I study whether social context affects the value-related content that appears in language model outputs and how reliably another model can identify it. To this end, I introduce a Generator-Inquisitor framework, in which one instance of a model is instructed to generate text that reflects a target value and another instance infers the underlying value from that text. I ground the framework in two established and cross-culturally validated psychological theories. The first is Schwartz's Theory of Basic Values, which organizes values into four higher-order dimensions---Openness to Change, Conservation, Self-Enhancement, and Self-Transcendence. The second is Fiske's Relational Models Theory, which distinguishes four fundamental relational contexts---Communal Sharing, Equality Matching, Authority Ranking, and Market Pricing. Crossing these yields a fully factorial $4 \times 4$ design, which I use to evaluate six state-of-the-art LLMs (GPT-4o, GPT-5-Mini, GPT-5, Gemini 2.0, Mistral Large 2512, and Llama 3.3 70B) across five experiments. Value recognition accuracy was generally high, well above the 25\% chance level, but varied systematically across value dimensions and relational contexts. Self-Enhancement was consistently the value dimension with the lowest recognition accuracy across all six models, with the difficulty most pronounced in Communal Sharing contexts. Valence scoring of generated texts revealed higher positivity for Self-Enhancement outputs compared to other value dimensions. Cross-model evaluation showed that value signals transferred reliably between models from different families (i.e., models from distinct developers and architectures), and a perturbation experiment showed that accurate classification depended on value-relevant content rather than superficial lexical cues. Together, these results show that the value-related content in LLM outputs, and how reliably it can be recognized, vary systematically with relational context, suggesting that evaluation of value alignment should be sensitive to the social context rather than treating values as fixed model attributes.

Degree

thesis:*
Grantor dc:publisher
University of Waterloo
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Ghorbansabagh, Maryam

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/10012/23968

Chain of custody

source
Harvested from
University of Waterloo
Base URL
dspacemainprd01.lib.uwaterloo.ca/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Ghorbansabagh, Maryam. How social context shapes value-related content in Large Language Model outputs. University of Waterloo, 2026. https://hdl.handle.net/10012/23968