I wonder if this is an unexpected consequence of training on unproblematic data? I could see politely phrased responses having a statistical bias towards validating others’ statements, rather than disagreeing.