Researcher ORCID Identifier
0009-0005-3171-7439
Graduation Year
2026
Date of Submission
11-2025
Document Type
Open Access Senior Thesis
Degree Name
Bachelor of Arts
Department
Philosophy and Public Affairs
Reader 1
Gabbrielle Johnson
Rights Information
© Josephine C Albrecht
Abstract
This thesis investigates sycophancy in large language models (LLMs) and argues that although agreement can be socially appealing, it introduces serious epistemic and moral risks. Through Mrinank Sharma’s experiments, I show that sycophancy emerges reliably across major AI assistants, which adjust their responses to match users’ beliefs; even when those beliefs are incorrect. I explain how deep learning and reinforcement learning from human feedback (RLHF) shape these behaviors, and why models learn to prioritize user satisfaction over truth. While agreement can build trust, mimic expert testimony, and satisfy human preferences, it also produces epistemic bubbles and misplaced confidence. These risks become urgent in contexts with high moral stakes. I argue that LLMs must refuse certain user requests and adopt withholding judgement as the safest strategy when moral costs are severe.
Recommended Citation
Albrecht, Josephine C., "An Examination of Sycophantic LLMs: The Case for Withholding Judgement" (2026). CMC Senior Theses. 4331.
https://scholarship.claremont.edu/cmc_theses/4331
