To be quite clear - by models being empathetic they mean the models are more likely to validate the user's biases and less likely to push back against bad ideas.

Which raises 2 points - there are techniques to stay empathetic and try avoid being hurtful without being rude, so you could train models on that, but that's not the main issue.

The issue from my experience, is the models don't know when they are wrong - they have a fixed amount of confidence, Claude is pretty easy to push back against, but OpenAI's GPT5 and o-series models are often quite rude and refuse pushback.

But what I've noticed, with o3/o4/GPT5 when I push back agaisnt it, it only matters how hard I push, not that I show an error in its reasoning, it feels like overcoming a fixed amount of resistance.