RLHF biases the model against negativity and in favor of this vacuous saccharine HR-creature tone that is reminiscent of Linkedin. It's not an inherent property of LLMs.