Of course any human can distinguish between sarcasm and irony.
It's an apt point. There are any number of studies showing how human judges disagree on tasks that people complain AIs fail at. This is certainly true in Natural Language Processing, for example, where it turns out that you can't get even a panel of expert judges to agree on assigning semantic parses (using e.g. Rhetorical Structure Theory) to complex statements, for example.
Andrew's test is just another example of the whole failed category of Imitation Game tests. The point of the IG, and Turing's entire essay, is not that we should use IG contests or other ill-defined "sure, humans can do X" acid tests to evaluate the state of AIs. It's a philosophical argument, essentially staking out a position congruent with American pragmatism (and thus rejecting metaphysics and the chancy bits of epistemology1): intelligence is what intelligence does.
The other problem with Andrew's test is that it makes AI into some singular, monolithic, all-or-nothing quality: either the machine is equivalent to some (again ill-defined) ideal person, or it's nothing at all. While it's useful to point out the many, many ways in which Google's big-data-and-deep-learning hammer fails to hit all the nails, much less deal with the screws, of human language, this business of "it can't do X so it means nothing" is not productive.
And to claim, as Andrew does, that there haven't been "any serious breakthroughs" in AI "in recent years" is just stupid. Maybe not as stupid as "smart" chat clients, but stupid nonetheless.
1I.e., all of epistemology.