Large earthquakes rupture faults over hundreds of kilometers within minutes. Finite-fault models image these processes and provide observational constraints for understanding earthquake physics. However, finite-fault inversions are subject to non-uniqueness and uncertainties. The diverse range of published models for the well-recorded 2011 $M_w$~9.0 Tohoku-Oki earthquake illustrates this issue, and details of its rupture process remain under debate. Here, we comprehensively compare 32 finite-fault models of the Tohoku-Oki earthquake and analyze the sensitivity of four commonly-used observational data types (geodetic, teleseismic, regional seismic-geodetic, and tsunami) to their slip features. We first project all models to a realistic megathrust geometry and a 1-km subfault size. At this scale, we observe low correlation among the models, irrespective of the data type. However, model agreement improves significantly with increasing subfault sizes, implying that their differences primarily stem from small-scale features. We then forward-compute geodetic and seismic synthetics and compare them with observations available during the earthquake. We find that seismic observations are sensitive to rupture propagation, such as the peak-slip rise time. However, neither teleseismic, regional seismic, nor geodetic observations are sensitive to spatial slip features smaller than 64~km. In distinction, the seafloor deformation predicted by all models exhibits poor correlation, indicating sensitivity to small-scale slip features. Our findings suggest that fine-scale slip features cannot be unambiguously resolved by remote or sparse observations, such as the four data types tested in this study. However, better resolution may become achievable from dense offshore instrumentation.