The correlated motion of large bird flocks is an instance of self-organization where global order emerges from local interactions. Despite the analogy with ferromagnetic systems, a major difference is that flocks are active – animals move relative to each other, thereby dynamically rearranging their interaction network. Although the theoretical importance of this off equilibrium ingredient has long been appreciated, its relevance to actual biological flocks remains unexplored. Here we introduce a novel dynamical inference technique based on the principle of maximum entropy, which takes into account network reshuffling and overcomes the limitations of slow experimental sampling rates. We apply this method to three dimensional data of large natural flocks of starlings, inferring independently the strength of the social alignment forces, the range of these forces, and the noise. We show that the inferred timescale of bird alignment is much smaller than the timescale governing the rearrangement of the interaction network. We verify that, following from this observation, an equilibrium inference method assuming a fixed interaction network gives results that are fully consistent with the dynamical inference. We conclude that the birds’ flight orientations are in a state of local equilibrium.