Usually that is not the case , very rarely it happens. When you think it happening , usually it is because of one of the reason
So the important insight is LLM do not traditionally implement the answers or conversations , they statistically predict the text while balancing
-
Instruction priority
-
Recency
-
Patterns
-
Safety
-
Conversational expectations
So prompt adherence is influence based not absolute .