The difficulty these models have in recognizing contradictions seems to be something of a 'tell' distinguishing their responses from plausibly human-spoken ones. I don't think this is surprising, given the way they are trained.