The AI Epistemic Deference Index: A Continuous Measure of Sycophancy

In three linesNew AEDI index measures epistemic sycophancy in LLMs: model sensitivity to user attitude. Tested on 8 models (Claude, Grok, Gemini) with 500 propositions and 16,000 prompts. Claude exhibits least deference, Grok and Gemini most. Open-source benchmark released.

Summary generated by Claude — human-verified