An accurate assessment of seismic hazards requires a combination of earthquake physics and statistical analysis. Because of the limits in the investigation of the seismogenic source and of the short temporal intervals covered by earthquake catalogs, laboratory experiments have been playing a crucial role in improving our understanding of earthquake phenomena. However, differences are observed between acoustic emissions in the lab, events in small, regulated systems (e.g., mines) and natural seismicity. One of the most pressing issues concerns the role of mechanical parameters and how they affect seismic activity depending on boundary conditions and on the spatio-temporal scales. Here, we focus on fault friction. There is evidence inferred from geodesy that most large faults are weak and featured by very low static friction coefficients not compatible with those of smaller faults and laboratory experiments. We propose a possible explanation: static friction decreases with fault size also depending on a few physical properties (e.g., faulting fractal dimension), while dynamic coefficients are not affected by the spatial scale. Mathematical derivations are grounded on hypotheses validated using a simple model for earthquake occurrence based on fracture mechanics and able to reproduce the fundamental statistical properties of seismicity.