Its hardly an original question. And the answer (we don’t know) isn’t original either. In case you were wondering, this is Overland and Wang, GRL 2013, doi: 10.1002/grl.50316 (PDF courtesy of V). Different but not entirely different to A sea ice free summer Arctic within 30 years?, also in GRL; or even A sea ice free summer Arctic within 30 years-an update from CMIP5 models by Wang and Overland.

W+O discuss three methods for predicting future Arctic sea ice: extrapolation, wild guesswork, and cherrypicking your favourite model (of course they don’t use those names, they’ve made up some fancier ones, but it comes to the same thing).
Extrapolation is to be done from ice volume, not ice extent. No expense is spared to fully justify this important choice: Their main points are that sea ice volume is decreasing at a rate that is
faster than sea ice extent, and that volume is a better variable than extent to use for sea ice loss. Well, you can’t argue with that, so I won’t try to.
Wild guesswork is the idea that ice loss occurs primarily through big loss years like 2007 and 2012, so we sort-of think of a trend, and then guess that things like 2007 or 2012 might happen, errm, sometimes.
Cherrypicking your favourite model is when you look at the sea ice observations (oddly, at this point you silently switch back to extent from volume, but since you do this silently you don’t have to explain why) plotted on top of spaghetti diagrams from CMIP and realise that you can make nothing of them. So perhaps you should throw the slower ones away.
These three different methods produce completely different answers, and so you end up concluding that you know nothing.
How might we do things better? Weeeeeellll, I’m not in the game anymore, so I can have fun. The O+W approach reminds me of Lord Dorwin, the aristocrat in Foundation whose idea of research is only to re-read the works of others. I’ve always thought (though I may not have said it very loudly, so please don’t challenge me to find me saying it somewhere) that plotting all the CMIP models is silly: some of them aren’t very good (I have a paper sort-of saying this: An Antarctic assessment of IPCC AR4 coupled models). Fig 3, reproduced above, needs to have the goo-n-dribble models removed (ah yes, cherrypicking: but no, it needs to be done on some objective grounds; see my paper). The most obvious thing about fig 3 isn’t that the obs disagree with the models, but that the spread in the models is enormous. But even with the junk removed I fear you’d find the obs retreat faster than the models; and I’m beginning to wear thin the idea that this is just a few years anomaly. So, really, we need better models and a better understanding of what is going wrong with the current models. That won’t come from spaghetti graphs or equation-free papers, though.