We consider the Glauber dynamics for the 2D Ising model in a box of side L, at inverse temperature β and random boundary conditions τ whose distribution P either stochastically dominates the extremal plus phase (hence the quotation marks in the title) or is stochastically dominated by the extremal minus phase. A particular case is when P is concentrated on the homogeneous configuration identically equal to + (equal to −). For β large enough we show that for any ε > 0 there exists c = c(β, ε) such that the corresponding mixing time Tmix satisfies limL→∞ P (Tmix ≥ exp(cL ε )) = 0. In the non-random case τ ≡ + (or τ ≡ −), this implies that Tmix ≤ exp(cL ε ). The same bound holds when the boundary conditions are all + on three sides and all − on the remaining one. The result, although still very far from the expected Lifshitz behavior Tmix = O(L 2 ), considerably improves upon the previous known estimates of the form Tmix ≤ exp(cL 1 2 +ε ). The techniques are based on induction over length scales, combined with a judicious use of the so-called "censoring inequality" of Y. Peres and P. Winkler, which in a sense allows us to guide the dynamics to its equilibrium measure.2000 Mathematics Subject Classification: 60K35, 82C20 and the right-hand side is an increasing function of σ; in accord with the notations of Section 1.2, L + J denotes the generator of the dynamics in J with + boundary conditions on ∂J (its invariant measure is of course π + J ) and L is the generator of the infinite-volume dynamics. One has then (using once more monotonicity)f 2 (4.23) which, together with (4.21), gives ρ(t) = Var + ∞ e tL f ≤ Var π + J e tL + J f + c e −c . (4.24) By (1.6), one has that Var π + J e tL + J f ≤ Var π + J