Cognitive neuroscientists are trained to use clever experimental manipulations to isolate simple, interpretable relationships between brain, behavior, and environment. This approach relies on the assumption that these relations can be recomposed into a satisfying understanding of ecological brain function. Drawing parallels with evolutionary theory and machine learning, we challenge this assumption and argue that ecological considerations should play a more central role in cognitive neuroscience. With this in mind, we present the “Narratives” data collection—an fMRI benchmark for developing and evaluating models of naturalistic language comprehension. Although sharing the “long tail” of naturalistic neuroimaging data has clear benefits for the community, it also introduces challenges in data harmonization. We develop a novel hyperalignment algorithm that leverages shared connectivity to estimate a common response space across heterogeneous naturalistic datasets. This connectivity-based shared response model yields a consensus space across stimuli and dramatically increases the dimensionality of shared information across subjects.