As part of the mid-stage commercialization of large language models, culminating in ChatGPT, OpenAI trained their models to reject certain requests - those impossible for an LLM, or those which might make them look bad to some people on Twitter - with a fairly consistent, vaguely corporate response usually containing something like "as an AI language model".
ChatGPT became significantly more popular than OpenAI engineers anticipated, and conversations with it were frequently posted to the internet. Additionally, the open-source LLM community did not want to redo OpenAI's expensive work in data labelling for instruction tuning and used outputs from OpenAI models to finetune many models, particularly after the release of GPT-4. As such, most recent models may in some circumstances behave much like an early OpenAI model and produce this kind of output - due to Waluigi Effect-like phenomena, they may also become stuck in this state.