Arctic sea ice is currently undergoing a period of rapid decline and concurrent high variability. Future projection uncertainty is dominated by internal variability in the short- and medium-term. In this thesis we use a large number of climate model simulations, which together can explain drivers and characteristics of variability, whilst diagnosing model limitations when compared with observations. We find that large ensembles from the Coupled Model Intercomparison Project - Phase 5 (CMIP5) generally simulate interannual variability within observational uncertainty. However, model realism is only achieved due to large observational uncertainties and spatially heterogeneous model biases. To reduce uncertainty of low-frequency variability, we find that climate modes of variability can be used skillfully to explain regional Arctic sea ice variability. The primary drivers of a decadal decline in Arctic sea ice are above trend global temperatures, a negative Interdecadal Pacific Oscillation (IPO) and a positive Atlantic Meridional Oscillation (AMO), which currently are aligned as such. The IPO across CMIP6 climate models is highly important for up to 15 years, but with large variations in magnitude and sign between models and realizations. With further investigation, we find the decadal predictive skill of the IPO is often dependent on the modulation by the slowly-changing Pacific-Arctic (PARC) atmospheric teleconnection. Considering the current negative phases of the IPO and PARC, an increased likelihood of an accelerated transition to an ice-free Arctic is expected by most climate models.