It's due to how the RLHF and instruction tuning was done. IIRC, even the builtin system prompt works this way in ChatGPT.