For many observers, UK Chief Scientist John Beddington’s argument that the world faced a ‘perfect storm’ of global events by 20301 has now become a prescient warning. Recent mention of ‘ghastly futures’2, ‘widespread ecosystem collapse’3 and ‘domino effects on sustainability goals’4 tap into a growing consensus within some scientific communities that the Earth is rapidly destabilizing through ‘cascades of collapse’5. Some6 even speculate on ‘end-of-world’ scenarios involving transgressing planetary boundaries (climate, freshwater and ocean acidification), accelerating reinforcing (positive) feedback mechanisms and multiplicative stresses. Prudent risk management clearly requires consideration of the factors that may lead to these bad-to-worst-case scenarios7. Put simply, the choices we make about ecosystems and landscape management can accelerate change unexpectedly.

The potential for rapid destabilization of Earth’s ecosystems is, in part, supported by observational evidence for increasing rates of change in key drivers and interactions between systems at the global scale (Supplementary Introduction). For example, despite decreases in global birth rates and increases in renewable energy generation, the general trends of population, greenhouse gas concentrations and economic drivers (such as gross domestic product) are upwards8,9—often with acceleration through the twentieth and twenty-first centuries. Similar non-stationary trends for ecosystem degradation10 imply that unstable subsystems are common. Furthermore, there is strong evidence globally for the increased frequency and magnitude of erratic events, such as heatwaves and precipitation extremes11. Examples include the sequence of European summer droughts since 201512, fire-promoting phases of the tropical Pacific and Indian ocean variability13 and regional flooding11, already implicated in reduced crop yields14 and increased fatalities and normalized financial costs9.

The increased frequency and magnitude of erratic events is expected to continue throughout the twenty-first century. The Intergovernmental Panel on Climate Change (IPCC) Sixth Assessment Report concludes that ‘multiple climate hazards will occur simultaneously, and multiple climatic and non-climatic risks will interact, resulting in compounding overall risk and risks cascading across sectors and regions’11. Overall, global warming will increase the frequency of unprecedented extreme events11, raise the probability of compound events15 and ultimately could combine to make multiple system failures more likely16. For example, there is a risk that many tipping points can be triggered within the Paris Agreement range of 1.5 to 2 °C warming, including collapse of the Greenland and West Antarctic ice sheets, die-off of low-latitude coral reefs and widespread abrupt permafrost thaw17. These tipping points are contentious and with low likelihood in absolute terms but with potentially large impacts should they occur. In evaluating models of real-world systems, we therefore need to be careful that we capture complex feedback networks and the effects of multiple drivers of change that may act either antagonistically or synergistically18,19,20. Prompted by these ideas and findings, we use computer simulation models based on four real-world ecosystems to explore how the impacts of multiple growing stresses from human activities, global warming and more interactions between systems could shorten the time left before some of the world’s ecosystems may collapse.

Intuitively, stronger interactions between systems may be expected to increase the numbers of drivers of any one system, change driver behaviour and generate more system noise. As a result, we would anticipate that higher levels of stress, more drivers and noise may bring forward threshold-dependent changes more quickly. For any particular system (for example, the Amazon forest) it is possible to envisage a time sequence that starts with one main driver (for example, deforestation), then multiple drivers (for example, deforestation plus global warming), more noise through extreme events (for example, more droughts and wildfires), with additional feedback mechanisms that enhance the drivers (for example, diminished internal water cycle and more severe droughts). A vortex could therefore emerge, with drivers generating noisier systems as climate variability and the incidence of extreme events increases. Under worst-case scenarios, the circle becomes faster as reinforcing feedbacks accelerate connections or human activities increase stress levels. However, extreme events could also counteract each other (for example, extreme droughts and extreme rainfall events) and interconnections could also have weakening effects (for example, where increased plant growth driven by increased CO2 is counterbalanced by increased temperatures and droughts. To date, there is limited observational evidence showing that ecosystems have a record of tipping between alternate stable states21.

Others19 offer a mathematical tripartite classification of critical transitions that includes slow driver bifurcations, rate-induced (fast/cumulative driver) and noise-induced (extreme event) tipping points. However, previous studies tend to focus on each of these categories individually. For example, there is a well-established body of physics and mathematical theory on ‘mean exit times’22, with studies investigating the timing of tipping points in rate-induced18,19,20 or noisy19,23,24 systems. However, despite calls for more experimental evidence of the impacts of climate variability and extremes on ecosystems25,26, the relative importance or combined effect of fast drivers, multiple drivers and noisy system drivers on the collapse of real-world ecosystems is not known. Critical transitions driven by current pollution forcings such as greenhouse gas emissions27 and nutrient loadings28 are likely to be new, well beyond the envelope of natural variability. Hence, we avoid the use of the terms critical transition and tipping points, used formally in dynamical systems theory to represent shifts to alternative attractors and focus on abrupt threshold-dependent changes (ATDCs) that would be perceived by society as the quantitative (for example, fish and stock integrity) and/or qualitative (for example, ecosystem functions) collapse of a desirable system state29,30.

We have selected a range of system dynamic models that have been previously used to demonstrate generalizable findings (for example, with regard to safely overshooting ATDCs27) and can be externally manipulated to simulate internal emergent ATDCs at local and regional scales—as if they were impacted through stronger connections to other systems. Reflecting modern ecosystems, these models show varied anthropogenic interactions, ranging from social-ecological systems with strongly coupled human–nature feedbacks to ecological systems with predominantly one-way interactions where ecosystems are influenced by the external impacts of people. The ability of these models to capture feedback loops, delays and interactions between components is well established31,32 and has motivated their use in various recent studies of sustainability and resilience21,33,34,35. Therefore, guided by the ref. 19 typology of tipping points, we aim to generalize the dynamics of increasing the numbers of drivers, their rates and variability (as proxies for stronger interactions between systems and noise) on the speed at which ATDCs are reached in four ecosystem dynamics models (Fig. 1): Lake Chilika lagoon fishery21,33, Easter Island36, Lake Phosphorus28,37 and a modified version of The Hadley Centre Dynamic Global Vegetation Model (TRIFFID) of forest dieback27,38.