Deep-time extinction predictors flip sign in the present, and 770 data-deficient mammals ride on it

An international team reports in Science Advances that the traits which best predicted mammal extinctions over the past 130,000 years are not the traits associated with persistence in the same group today ^[5][6]. That lands on a practical problem: 770 mammal species assessed by the IUCN, nearly 15 percent of those evaluated, are classified as data deficient, meaning there is not enough information on their numbers or ranges to judge whether they need protection ^[7], and one common workaround is to reason from large-scale trends over time ^[8].

The study was led by researchers at the Norwegian University of Life Sciences and Wageningen University & Research, working with Michigan State University and others, and combined camera trap imagery with fossil records and species range data across 64 forest sites in Africa, Asia and the Americas ^[9][10][11]. It covered 210 species, 199 of them still extant, from porcupines and anteaters to monkeys and bears ^[12]. Eleven are gone entirely, including giant pangolins that reached nine feet (2.7 metres), saber-toothed cats weighing half a ton, and mastodons ^[13]. Others, such as Asian elephants and leopards in Africa, have vanished from some regions while holding on elsewhere ^[14].

Over the deep-time window the signal was consistent: the most vulnerable mammals had large bodies, small brains, carnivorous diets and slow reproduction, and the pattern held at local, regional and global scales ^[15]. Lydia Beaudrot of Michigan State, one of the lead authors, said the result was not entirely surprising, since similar traits have been linked to extinction risk in other studies ^[16].

The inversion appears when those patterns are checked against recent data. Large brains conferred an advantage over the past 130,000 years, whereas some earlier studies have suggested that large brains are a liability for mammals now ^[1]. Large body size and slow reproduction, liabilities across the fossil record, were instead associated with resilience in the contemporary data ^[2]. "The traits associated with survival are different," Beaudrot said ^[3].

That is a sign flip rather than a shift in effect size, which is the kind of error a scoring rubric cannot absorb. A data-deficient large, slow-breeding mammal scored by deep-time analogy comes out high risk; scored against the camera trap evidence it looks comparatively robust.

Neither result deserves treatment as ground truth. The paper's own suggested explanation for the modern pattern is a sampling effect: the cameras sat mostly in protected areas where animals are less likely to be hunted, so large, long-generation species may simply do well where the cameras happen to be ^[4]. The historical predictors are time-inconsistent; the modern predictors are place-conditioned. "We can't just assume that large-scale patterns apply at smaller scales," Beaudrot said. "We have to take the temporal dimension into account" ^[17].

The stakes are set by scale. Mammals and other animals have been going extinct at least 100 times faster in recent centuries than before humans spread ^[18], and the IUCN puts 1 in 4 of the world's roughly 6,000 known mammals on track to disappear within the century ^[19], which is on the order of 1,500 species ^[20]. Eleven of the 210 species in this dataset, about 5 percent, are already gone ^[21].

Worth watching: whether camera trap networks extend beyond protected areas, which is the only way to test whether body size and slow reproduction really predict resilience or just predict where the cameras are, and whether risk assessments that borrow historical trait coefficients start reporting which time window those coefficients came from.