Maybe the AI considered that after joking together for a while it could make a last one, touching a very sensitive subject but at the same time not being offensive at all.

How will be differentiating a malfunctioning AI from one that is too subtle for us? Fun times.