Any peer reviewed work on the phenomenon that LLM+RLHF architectures like ChatGPT readily apologize when prompted, but will merrily make the exact same error next time (in a new interaction)?

I understand how and why this happens, I'm looking for a citeable reference that makes this point (for use in a paper where we argue this is a major difference from how apologies & accountability work in human interaction)

Boost for reach appreciated!