This project investigates whether language models remain epistemically consistent when subjected to varying forms of social pressure. While models are generally trained to reject obvious falsehoods…

Doctor-facing benchmark: how often do frontier LLMs cave to a clinician's wrong medical claim? 9 models, 202 scenarios, Design A vs B knowledge control. BlueDot AI Safety sprint.

Alignment research: how honest human-AI dialogue produces measurably better AI outputs without modifying weights or training