LLM chat window screenshot
I have seen this more than once in LLM models. Recently started experimenting with llama2.
I observe that the model gives a results and then later realises that it might have been inappropriate, and then corrects itself rather than not giving incorrect/ inappropriate answer in the first place.

What is this phenomenon called, why does it happen and how could it be fixed?

Had set model temperature to 1, and gave a brief intro mentioning my name as description under SYSTEM.