Confidence in scientific models accumulates by continuously validating the model’s predictions by observations. We compared the seismic‐hazard forecasts of the four published versions of the U.S. Geological Survey National Seismic Hazard Maps with observed ground motions. A large dataset is necessary for a statistically meaningful comparison, and so our comparison was based on an aggregated approach such that the observations and the forecast in a region (California, and the central and eastern United States [CEUS]) were combined and compared as a whole. We used instrumental records in California and macroseismic intensity in the CEUS since 2000 as the observation, which was largely prospective to the hazard maps. We verified that the observed seismic hazard based on macroseismic intensity was consistent with that based on instrumental records, making model evaluation in the CEUS, for which instrumental records were almost nonexistent, viable. The observed hazard was found to be generally consistent with the forecasted one for peak ground acceleration (PGA) in California and for both PGA and spectral acceleration at 1 s (SA1) in the CEUS. Forecasted hazard for SA1 for California appeared to be conservative. Recent versions of the hazard map were in better agreement with observations in California. Small earthquakes, as expected, were found to have insignificant impact on SA1. Induced earthquakes in the CEUS have increased the observed seismic hazard but did not invalidate the hazard model as a whole. We examined the resolving power of the test by computing its statistical power.