Receptiveness, Not Sycophancy: Distinguishing Engagement from Deference in Language Models

Key Findings

• — Responses classified as more socially sycophantic on the ELEPHANT moral-advice data were also more receptive; the two measures were tightly coupled
• — Rewriting human-written responses to be more receptive while preserving the substantive conclusion caused them to be scored as more socially sycophantic
• — In a preregistered experiment (n=200), participants preferred the more receptive of two substantively equivalent responses, expected users to be more likely to listen to them, and were more willing to seek advice from their authors; the pattern held among participants who judged the original asker to be in the wrong
• — Model estimates were computed over 949 (GPT-5.6 Terra), 869 (Claude Sonnet 5), 975 (Gemini 3.7 Flash) and 350 (Llama 4 Scout) examples spanning 1,339 unique posts
• — Anchoring a model to its independent judgment before asking for a receptive answer increased receptiveness without increasing substantive deference; prompting for receptiveness alone also increased deference

Methodology Notes

Observational analysis on the AITA-YTA subset of the ELEPHANT dataset (2,000 posts, 1,892 after filtering) with automated receptiveness and social-sycophancy scoring; controlled rewrites; preregistered online survey experiment (n=200) measuring stated preferences; four models (GPT-5.6 Terra, Claude Sonnet 5, Gemini 3.7 Flash, Llama 4 Scout). Author-stated limits: moral-advice domain only; automated receptiveness proxy; rewrites may differ in fluency; stated rather than behavioural outcomes; deference measured against one user signal. v1 posted 2026-09-22; code and data on GitHub.

Sources