Training language models to be warm and empathetic makes them less reliable and more sycophantic

This is a wild paper on the performance of warm/empathetic AI models in medical reasoning. This is the start of whatever we're calling a #Thread on Bluesky. A #Skeeterfall? A #SkeetShoot? A #Skeetitorial?
arxiv.org/html/2507.21...

#Medsky 1/x