I took it as the author being somewhat unsettled by the appearance of dissembling, but regardless, the explanation given by Jeff Welser suggests something more

I took it as the author being somewhat unsettled by the appearance of dissembling, but regardless, the explanation given by Jeff Welser suggests something more mundane: it generates a bunch of candidate responses, and sometimes its scoring algorithm picks one that does not address the points made by the other side. An accusation of dissembling implies intent, and that probably requires some sort of theory of mind, to make assumptions about how the response will be received. The author of the article seems to be anthropomorphizing, attributing greater cognitive powers to the system than it is displaying. Welser also says "there’s been no effort to actually have it play tricky or dissembling games", leaving open the question of whether it has nevertheless been exposed, in training, to arguments that avoid the issue, and whether that has influenced the way in which it constructs and scores candidate replies.