The hypothesis that destructive mass extinctions enable creative evolutionary radiations (creative destruction) is central to classic concepts of macroevolution1,2. However, the relative impacts of extinction and radiation on the co-occurrence of species have not been directly quantitatively compared across the Phanerozoic eon. Here we apply machine learning to generate a spatial embedding (multidimensional ordination) of the temporal co-occurrence structure of the Phanerozoic fossil record, covering 1,273,254 occurrences in the Paleobiology Database for 171,231 embedded species. This facilitates the simultaneous comparison of macroevolutionary disruptions, using measures independent of secular diversity trends. Among the 5% most significant periods of disruption, we identify the 'big five' mass extinction events2, seven additional mass extinctions, two combined mass extinction-radiation events and 15 mass radiations. In contrast to narratives that emphasize post-extinction radiations1,3, we find that the proportionally most comparable mass radiations and extinctions (such as the Cambrian explosion and the end-Permian mass extinction) are typically decoupled in time, refuting any direct causal relationship between them. Moreover, in addition to extinctions4, evolutionary radiations themselves cause evolutionary decay (modelled co-occurrence probability and shared fraction of species between times approaching zero), a concept that we describe as destructive creation. A direct test of the time to over-threshold macroevolutionary decay4 (shared fraction of species between two times ≤ 0.1), counted by the decay clock, reveals saw-toothed fluctuations around a Phanerozoic mean of 18.6 million years. As the Quaternary period began at a below-average decay-clock time of 11 million years, modern extinctions further increase life's decay-clock debt.