LLMs do a bad job noticing "not".

So if you have a long conversation where you say "not X" over time it has a high likelihood of thinking just "x."