Several studies have shown that local magnitude, ML, and moment magnitude, M, scale differently for small earthquakes (M < ~2) than for moderate to large earthquakes. Consequently, frequency‐magnitude relations based on one or the other magnitude type cannot obey a power law with a single exponent over the entire magnitude range. Since this has serious consequences for seismic hazard assessments, it is important to establish for which magnitude type the assumption of a constant exponent is valid and for which it is not. Based on independently determined M, ML and duration magnitude, Md, estimates for 5,304 events near Parkfield, we confirm the theoretically expected difference in scaling between the magnitude types, and we show that the frequency‐magnitude distribution based on M and Md follows a Gutenberg‐Richter relation with a constant slope, whereas for ML it is bilinear. Thus, seismic hazard estimates based on ML of small earthquakes are likely to overestimate the occurrence probability of large earthquakes.