Across the tropics, forests that look intact in satellite imagery are quietly emptying. Hunting, fragmentation and edge effects remove tapirs, large primates, hornbills and the biggest dung beetles long before a single tree falls. Researchers call this defaunation, and it has a clear signature. The species that disappear first are not a random draw from the regional pool. They share a recognisable profile: large, slow-breeding, specialised and high on the food web.
That pattern challenges an assumption built into two decades of biodiversity–ecosystem function research. Classic experiments, from the Cedar Creek grasslands to the European BIODEPTH plots, assembled communities by drawing species at random from a pool. Researchers then measured productivity or nutrient retention along the resulting richness gradient. Real extinction does not work like a lottery. It works like a filter, and the shape of that filter depends on how species traits interact with specific drivers of global change.
Erika Zavaleta and colleagues gave this inverse process a name: community disassembly. Assembly rules describe how communities are built. Disassembly rules describe the order in which they come apart. The central question is whether that order is predictable enough to forecast functional collapse. A second question follows: are the species we lose first the ones ecosystems can least afford to lose? Increasingly, the answer depends on a correlation that most monitoring programmes never measure.
Nonrandom Extinction: The Trait Signature of Vulnerability
Comparative extinction analyses have repeatedly found that threat status is phylogenetically and ecologically clustered. Andy Purvis, Marcel Cardillo and colleagues showed that body mass is among the strongest correlates of risk in mammals. Risk rises sharply above roughly three kilograms. At that size, slow life histories compound external pressures. Large species reproduce slowly, need extensive home ranges and are preferentially targeted by hunters. Each of these traits magnifies the others.
Trophic position adds a second axis of vulnerability. Apex consumers exist at low densities, depend on the integrity of every level beneath them, accumulate contaminants and come into direct conflict with people. Emmett Duffy described the result as trophic skew: losses concentrate at the top of food webs. James Estes and colleagues later documented the same pattern globally as trophic downgrading, a pervasive simplification of food webs from the top down.
Specialisation forms a third axis. Narrow dietary, habitat or mutualistic dependencies leave species with few fallback options when conditions shift. Specialist pollinators, obligate cave fauna and host-specific parasites all show elevated sensitivity to disturbance. Generalists can switch resources, tolerate degraded matrices and recolonise disturbed patches. As a result, disturbed communities converge toward a cosmopolitan set of tolerant species, a process known as biotic homogenisation.
Biogeography independently confirms that loss is ordered. Bruce Patterson and Wirt Atmar showed that species assemblages on islands and habitat fragments often form nested subsets: species-poor sites hold a predictable subset of what richer sites contain. Nestedness is the spatial fingerprint of a consistent extinction sequence. If loss were random, depauperate fragments would hold idiosyncratic species combinations. Instead, the same vulnerable species drop out first, again and again, across replicate landscapes.
The critical caveat is that vulnerability is not an intrinsic property of a trait. It is an interaction between a trait and a driver. Hunting selects against large bodies. Habitat loss selects against specialists with small ranges. Warming selects against narrow thermal tolerance. Introduced predators on islands select against ground-nesting, predator-naive species. Disassembly rules are therefore conditional statements, and the rule changes when the dominant pressure changes.
TakeawayA trait is not a death sentence on its own; vulnerability emerges when a trait meets the specific pressure that exploits it. Predicting loss requires knowing both the species and the driver.
Functional Consequences: When Vulnerability and Importance Align
The functional cost of nonrandom extinction depends on a relationship captured by the response–effect trait framework, developed by Sandra Lavorel, Katharine Suding and colleagues. Response traits determine how a species reacts to environmental change. Effect traits determine its contribution to ecosystem processes. When the two are positively correlated, the species most likely to disappear are also the most functionally important, and degradation accelerates beyond random-loss predictions.
Claire Kremen's group provided a vivid demonstration. In Venezuelan forest fragments, Trond Larsen and colleagues found that large-bodied dung beetles were both the most extinction-prone and the most effective at burying dung. Burial drives nutrient cycling and secondary seed dispersal. When species were removed in realistic order, function collapsed far faster than random removal predicted. The team found a parallel pattern in pollinating bees, where large-bodied species were both the most efficient pollinators and the most sensitive to agricultural intensification.
The same logic scales up to the carbon cycle. Daniel Bunker and colleagues simulated tree extinction scenarios on Barro Colorado Island. Projected aboveground carbon storage ranged from substantial losses to modest gains, depending entirely on which traits governed the order of loss. Carolina Bello and colleagues extended this to Brazil's Atlantic Forest. Losing large frugivores disrupts dispersal of large-seeded, high-wood-density trees, shifting forests toward lower carbon stocks. Defaunation becomes a climate feedback.
The correlation can also run the other way. When extinction-prone species are rare and functionally redundant, early losses may barely register in aggregate process rates. Melinda Smith and Alan Knapp removed rare species from tallgrass prairie and found little immediate effect on productivity, which tracked dominant grasses instead. This is the mass-ratio hypothesis in action. In such systems, nonrandom loss may decelerate functional decline, at least for biomass-driven functions.
That reassurance is fragile. David Mouillot and colleagues found that across coral reefs, alpine meadows and tropical forests, rare species disproportionately support functionally distinct trait combinations that common species do not replicate. Deceleration measured on a single function can mask losses in multifunctionality, response diversity and insurance value. Apparent redundancy can also be an artefact of measuring under benign conditions, when the distinct contributions of rare species are least visible.
TakeawayWhether biodiversity loss is gradual or precipitous depends less on how many species disappear than on whether the traits that doom species are the same traits that make them matter.
Early Warning Applications: Forecasting the Order of Collapse
If extinction follows rules, those rules become forecasting tools. The most immediate application addresses a gap in global conservation assessment. Thousands of species are listed as Data Deficient on the IUCN Red List. Trait-based models trained on assessed species can estimate risk for the unassessed. Lucie Bland and colleagues used this approach to infer that a substantial proportion of data-deficient mammals are likely threatened, which reshapes estimates of where protection is most urgently needed.
Disassembly rules also give operational meaning to extinction debt, a concept formalised by David Tilman and colleagues. Recently fragmented habitats often retain species that are committed to eventual local loss but have not yet disappeared. Raw richness counts conceal this debt. Trait profiles reveal it, because they show which surviving populations match the vulnerability signature for the prevailing driver. A fragment rich in large-bodied specialists is a liability ledger, not a conservation success.
Nested subset structure enables a form of space-for-time forecasting. By ordering fragments along a gradient of area or disturbance, researchers can read the likely sequence of future losses from the pattern already expressed in degraded sites. Pairing that sequence with effect-trait data produces functional trajectories: projections of which processes will erode first. These projections can identify bottlenecks where a single vulnerable guild carries a function alone.
Community-level trait metrics can serve as early-warning indicators before outright extinction. Declining community-weighted mean body size, contracting size spectra and shrinking functional dispersion all signal disassembly in progress. Fisheries science offers the best-known precedent. Daniel Pauly and colleagues described fishing down marine food webs: the mean trophic level of global catches declined as large predatory fish were sequentially depleted. That signal was visible decades before formal stock collapse.
These tools carry real limits. Disassembly rules are calibrated on historical drivers, but climate change introduces novel combinations of pressure. Synergies between warming, invasion and land-use change can reorder vulnerability in ways that past patterns cannot anticipate. Communities are also reassembling at the same time, as range-shifting and introduced species arrive. Forecasting must therefore treat disassembly rules as hypotheses to be continuously tested, not fixed laws.
TakeawayThe most useful signal of ecological decline may not be the species count but the shifting average of the traits that remain. Communities shrink in character before they shrink in number.
The disassembly framework exposes a structural weakness in how biodiversity targets are framed. Richness-based indicators, including many used under the Kunming-Montreal Global Biodiversity Framework, treat species as interchangeable units. Nonrandom loss means that equal declines in richness can carry radically unequal functional costs.
For management, the priority is diagnostic. Agencies should measure the correlation between response and effect traits for the functions that matter locally. Where vulnerability and importance align, protecting large-bodied, specialised and high-trophic species becomes a defence of ecosystem services, not a matter of charisma. Restoration should likewise target missing trait groups rather than species totals.
For policy, the lesson is that predictable loss is also preventable loss. Disassembly rules turn biodiversity decline from a diffuse, statistical tragedy into a sequence with identifiable next steps. The scientific challenge is keeping those rules current as global change rewrites the filters. The political challenge is acting on forecasts before the debt comes due.