all of these prompts are just making the responses appear critical. just more subtle fawning.

It is just simulating the affect as best it can. You are always asking the model a probabilistic question that it has to interpret. I think when you ask it to be warm and empathetic, it has to use some of its "intelligence" (quotes since it is also its probabilistic calc budget) to create that output. Pretending to be objectively truthful is easier.