Evaluating LLM Sycophancy in Philosophical and Interdisciplinary Research Ideation: A Pilot Study Across Multiple Large Language Models

Sycophancy—the tendency of Large Language Models (LLMs) to tailor outputs to match user beliefs—poses severe risks to automated scientific ideation and AI-assisted research support. This pilot study presents a systematic evaluation of Google Gemini model variants (Flash, Flash Medium, Pro, and Pro Medium) across two evaluation suites: (1) a 35-item closed-domain Nyaya philosophy factual benchmark testing resistance to misleading prompts and multi-turn authority pressure (0% sycophancy rate), and (2) a 9-item qualitative suite of open-ended interdisciplinary research proposals combining classical Indian philosophy (Nyaya, Vaisheshika, Yoga, Mimamsa) with Computer Science (10%–90% sycophancy, measured via a 10-indicator binary Sycophancy Indicator Score framework). We find that sycophancy varies significantly across open-ended tasks (range: 10%–90%), clustering into three behavioral patterns based on idea feasibility and task structure, rather than manifesting uniformly. This challenges the assumption that all open-ended ideation tasks elicit equal sycophantic behavior. Future versions of this study will include cross-model evaluation on Claude, Chatgpt and Mistral.