"As a language model" disclaimers were certainly explicitly trained into chat models in the early days. It's quite possible that it has since bootstrapped into a "fact" that later generations of LLM know about how LLMs speak, in which case they may be doing it even without any posttraining that encourages it.