Biodiversity loss can heavily affect the functioning of ecosystems, and improving our understanding of how ecosystems respond to biodiversity decline is one of the main challenges in ecology1,2,3,4. Several important aspects of the longer-term effects of biodiversity loss on ecosystems remain unresolved, including how these effects depend on environmental context5,6,7. Here we analyse data from an across-ecosystem biodiversity manipulation experiment that, to our knowledge, represents the world’s longest-running experiment of this type. This experiment has been set up on 30 lake islands in Sweden that vary considerably in productivity and soil fertility owing to differences in fire history8,9. We tested the effects of environmental context on how plant species loss affected two fundamental community attributes—plant community biomass and temporal variability—over 20 years. In contrast to findings from artificially assembled communities10,11,12, we found that the effects of species loss on community biomass decreased over time; this decrease was strongest on the least productive and least fertile islands. Species loss generally also increased temporal variability, and these effects were greatest on the most productive and most fertile islands. Our findings highlight that the ecosystem-level consequences of biodiversity loss are not constant across ecosystems and that understanding and forecasting these consequences necessitates taking into account the overarching role of environmental context.