I A Deer Dies in the Forest
A deer dies in the forest. What happens next?
Within days, the carcass begins to rot. Crows, foxes, and beetles take it apart, bit by bit, dragging the bones in different directions. A few months later, only scattered bones remain. Exposed to sun and rain, they begin to crack: first fine lines appear on the surface, then layers flake away along them, until the bones finally crumble.
Someone has watched this happen, year after year. Beginning in 1975, the American paleontologist Anna Behrensmeyer numbered the bones of dead wildebeest, zebras, and elephants on the grasslands of Kenya’s Amboseli National Park. Every few years she went back to check on them. She kept going back for more than forty years.
She divided bone weathering into six stages. On those grasslands, bones exposed at the surface usually fall apart into unrecognizable fragments within ten to fifteen years. Even on a single bone, the side turned toward the sun often weathers faster than the side against the ground; bones in the shade of a tree weather more slowly than those in the open. In reckoning a bone’s fate, you have to count even which side faces up.
A skeleton left on the surface therefore has a window of little more than a decade. This is the route taken by the vast majority of animals that die in forests and grasslands every year.
The seabed offers no sanctuary either. When a dead whale sinks into deep water, sharks, hagfish, and crabs strip off the flesh. Then crowds of small animals move onto its bones and live there for decades.
In 2002, the deep waters of California’s Monterey Bay yielded a kind of worm that specializes in eating whale bone. It has neither a mouth nor a digestive tract. Instead, it drives rootlike tissue into the bone, secretes acid to dissolve the mineral, and relies on bacteria inside its body to digest the collagen and fat. Fossils have never been the rule. Fossils are the exception. The same goes for us.
Demographers estimate that more than a hundred billion human beings have been born. Yet the fossils of humans and their ancestors, spanning millions of years and gathered from all over the world, represent only a few thousand individuals. Most amount to a tooth or a piece of jaw. This essay asks: by what rules are those exceptional few selected?
To become a fossil, the body must first be buried within that decade or so. Sediment carried by a flood, a landslide, a fall into a lake or river: any of these might do it. Most forest floors offer none of them.
The study of this succession of events is called taphonomy, a name coined in 1940 by the Soviet paleontologist Ivan Efremov. It means “the laws of burial.” Efremov dug for dinosaurs in Mongolia’s Gobi Desert and later became one of the Soviet Union’s best-known science-fiction writers, the author of the novel Andromeda.
That is only the first hurdle. Burial is followed by a whole series of others.
The early chemical environment determines whether the bone continues to dissolve or is replaced and secured by minerals. What is in the pore water? How acidic or alkaline is it? How much oxygen is available, and what are the surrounding microbes doing? These decide whether the bone dissolves away or stays. This stage runs on the brisk timetable of chemical reactions, often taking only a few years to a few decades.
Farther down, pressure flattens it. If mineralization happened early, its three-dimensional structure may survive. If not, it becomes a thin sheet. Deeper burial brings geothermal heat, which alters both its minerals and its organic remains.
Then comes the step that demands the greatest good fortune: it has to return to the surface. A fossil permanently buried ten kilometers underground might as well not exist for us. And staying buried is the fate of the overwhelming majority of sediments.
Once it reaches the surface, destruction starts again: oxidation, freezing and thawing, tree roots, running water. Every rock layer has a “collecting window,” wedged between “too deep to reach” and “exposed so long that it has already been destroyed.”
The final hurdle is no longer a natural process. Someone has to go there, recognize what it is, and put it on record. In 1996, a farmer in Beipiao, Liaoning, dug a small dinosaur fossil out of a hillside near his home. It had a fringe of fuzz around it and was later named Sinosauropteryx. That area became one of the world’s most important sources of feathered dinosaurs half thanks to lakes and volcanic ash more than a hundred million years ago, and half thanks to the thousands upon thousands of local hands turning over slabs of rock.
Is the outcrop beside a road? Can researchers enter the country? Does anyone specialize in this particular group of organisms? Is the specimen complete enough to be named? Every one of these helps decide whether it ever makes it into a database.
A 2022 survey of global paleontological research over recent decades found that researchers from high-income and upper-middle-income countries participated in producing about 97 percent of fossil data. Another study measured sampling against the area of sedimentary rock available for exposure and estimated that, outside North America and Europe, average sampling intensity was only about 45 percent of the level in those two regions.
Write this succession of hurdles as a chain and its most important property becomes clear. The probabilities multiply. Every link carries a probability below one, and all of them must be multiplied together. Each probability is conditional on having passed the earlier hurdles; the hurdles need not be independent.
If any one link approaches zero, no amount of success farther along can rescue it. Whether a group enters the record is often decided by its weakest link.
The second property matters even more: each link filters out something different. Decay removes the soft; transport removes the light; compaction removes the hollow; geothermal heat removes the organic; collecting removes the small and inconspicuous. The holes in each sieve have a different shape. What reaches the end is “the set that meets the requirements of every sieve at once.”
Take an example. A patch of seabed originally supports a hundred animal species, fifty with shells and fifty without. The fossils it leaves will not be ten species drawn at random from the hundred. They will be ten drawn from the fifty with shells.
Scale that sample up by any factor you like and you will never recover the original hundred. The fossil record is an inventory that has undergone a new round of selection. Random thinning can permit statistical estimates of the original abundance under suitable assumptions; scaling alone still cannot recover a complete species list. Select it anew and the remainder carries a systematic bias that no proportional enlargement can undo. Everything that follows in this essay starts here.
II Materials Are Not Equal
Even among hard shells, the odds of survival can differ by more than an order of magnitude. The clearest example involves two forms of calcium carbonate.
Calcium carbonate, the principal constituent of seashells, has two common crystalline forms. One is calcite, which is relatively stable. England’s dazzling white cliffs at Dover are made of countless calcite platelets only a few micrometers across. They were the coverings of single-celled marine algae, settling layer upon layer onto the seabed over more than thirty million years of the Late Cretaceous.
The other is aragonite, which readily dissolves under burial conditions or slowly transforms into calcite. Its name comes from Molina de Aragón in Spain, the locality where it was first described.
The chemical composition is the same. Change the arrangement and you change the fate. Pearls, and the iridescent mother-of-pearl lining a shell, are aragonite. An oyster’s shell is mostly calcite.
Most snail shells are aragonite. So are the skeletons of today’s reef-building corals. Brachiopod shells are mostly calcite; in Section IV, we will see just how deep an imprint that has left on their fossil record.
Which form an animal uses was settled long ago in the history of its lineage. The bill comes due in rock layers hundreds of millions of years later. In 2000, two British researchers found shell beds in rocks more than four hundred million years old on the Swedish island of Gotland. Silica had filled the shells very early, effectively “freezing” them in place.
They compared these with nearby rocks that lacked such protection. In the silicified shell beds, mollusks accounted for about 70 percent of species; in the unprotected rocks, only about 3 percent. Almost the entire contingent of aragonite-using mollusks had dissolved away.
The same animals, on the same stretch of seabed: some of their shells remained, while others were gone. In the fossil record, the ones that were gone might as well never have existed.
This changes which groups appear richer in species. It does more than subtract a few specimens. The extent of dissolution varies enormously from place to place, and no single conversion factor works across different ages and sedimentary basins.
Teeth, bones, and scales made of calcium phosphate are usually much more durable. A sensational debate once broke out over how long a bone’s original organic material can survive. In 2005, the American researcher Mary Schweitzer soaked a piece of a 68-million-year-old Tyrannosaurus rex leg bone in weak acid. When the minerals had dissolved, small, flexible tubes remained that looked like blood vessels.
Were they original tissues, or films left by bacteria that arrived later? The argument continues today. A shark replaces more than ten thousand teeth in its lifetime, which helps make shark teeth one of the commonest vertebrate fossils.
One ancient giant shark, estimated to have exceeded ten meters in length, had a name that meant “big tooth.” Today it is known almost entirely from teeth and a small number of vertebrae. No complete cartilaginous skeleton has survived.
Tooth enamel is the hardest substance in the body. Much of the study of fossil mammals is, in practice, the study of teeth. Many small mammals from the age of dinosaurs are still known only from a few teeth. Researchers use the arrangement of their cusps to infer whether they ate insects or plants.
Phosphate can also replicate soft tissue early in burial, but the conditions are exacting. To reproduce muscle, which decays with particular speed, mineralization often has to begin within days of death.
Silica can form an organism’s skeleton, or enter its tissues later. An experiment in the hot springs of Tateyama, Japan, left blocks of wood soaking in silica-rich water. Within seven years, silica accounted for almost 40 percent of their weight.
Under the right conditions, mineral fixation can proceed much faster than our usual picture of geological time suggests. In Arizona’s Petrified Forest National Park, tree trunks more than two hundred million years old lie across the desert, every growth ring turned to colored quartz. The same happens to bone: minerals often fill the pores of fossil dinosaur bones. The fine structure is still there, but the bones are much heavier than before.
The outer covering of an arthropod is another material again, made chiefly of chitin and proteins. Decay experiments show that some of its organic structure can be selectively retained, while the soft tissues inside disappear much faster.
An insect fossil, then, is often its outer casing, its muscles, internal organs, and nerves long gone. Many ancient insects have left nothing but a pair of wings. The giant dragonfly with a wingspan of about seventy centimeters that lived three hundred million years ago was first recognized from fossil wings found in a French coal mine.
Plants follow the same rule. Woody parts resist decay better than tender ones; waxes and cuticle on a leaf’s surface can leave an organic film. When you look at an ancient leaf, you see its outline, its stomata, and the patterns of its epidermal cells. The contents of those cells vanished long ago.
Plants possess something else that is almost indestructible: the outer walls of pollen grains and spores. They are made of an exceptionally stable polymer that can withstand even a bath in concentrated acid.
Geologists dissolve the rock away, and what remains is often thousands upon thousands of ancient pollen grains. They can reveal the age of a stretch of strata and the plants that grew nearby. Petroleum geologists learned long ago to use them to correlate rock layers from different boreholes.
III The Narrow Routes Open to Soft Bodies
A body without a hard shell can survive too, but only by taking one of a few narrow routes.
The first is carbonaceous compression. In 1909, Charles Walcott of the Smithsonian Institution discovered soft-bodied animals flattened into thin black films in the Burgess Shale of the Canadian Rockies. The outlines of tentacles, legs, and guts were all there.
In 1984, Hou Xianguang of the Chinese Academy of Sciences’ Nanjing Institute of Geology and Palaeontology split open a piece of mudstone at Maotianshan, in Chengjiang, Yunnan. Inside was a complete arthropod without a hard shell. The discovery opened up the Chengjiang biota, roughly ten million years older than the Burgess Shale; the site became a World Heritage Site in 2012. More than two hundred species have now been described there, many arriving on the scene already equipped with eyes and jointed legs.
Over his lifetime, Walcott collected more than sixty thousand specimens at the Burgess Shale. Many waited in drawers until decades after his death before anyone took them out for another close look. Carbonaceous compressions preserve overall outlines and the positions of internal organs in a plane particularly well, but almost all depth is lost. It is rather like looking at a photograph of a flattened specimen.
The second route is phosphatization. Southern Sweden has black limestone nodules that locals call “stinkstone”: break one open and a smell comes out. Researchers slowly dissolve this rock in acid and filter out tiny arthropods about a millimeter long, five hundred million years old. They are three-dimensional, with even the bristles on their legs intact. If phosphate precipitates early enough, it can replicate muscles and digestive tracts down to the cellular scale. But the window lasts only a few days after death.
The third is pyritization. In the late nineteenth century, shale beside a small creek in New York State yielded trilobites 450 million years old. Their bodies had been replaced by golden pyrite, preserving even legs and antennae that ordinarily disappear.
Charles Beecher at Yale spent years studying them. For the first time, people could see trilobite legs and antennae in such completeness. Pyritization requires iron, sulfur, microbes, and the decomposition of organic matter to converge in time and space. Yet if the pyrite later oxidizes, it can destroy the fossil in turn. What once saved it may ultimately undo it.
The Messel Pit near Frankfurt, Germany, offers another kind of compression. In the oil shale of a lake bottom 47 million years old lie little horses still carrying their unborn young, birds whose feather colors can be discerned, and the last meals left in stomachs. In the 1970s and 1980s, this abandoned mine nearly became a landfill. Local people and scientists fought for years; eventually the state government bought it. In 1995, it became a World Heritage Site.
Elsewhere in Germany, at Solnhofen in Bavaria, the limestone is so fine-grained that nineteenth-century printers used it for lithographic plates. Quarry workers peeled the slabs apart one by one. In 1861, one revealed the impression of a feather; soon afterward came Archaeopteryx. More than a century later, only a dozen or so Archaeopteryx specimens have emerged from the area.
The fourth route is silicification. Scotland’s Rhynie chert, discussed in Essay VI, sealed four-hundred-million-year-old plants, cells and all, inside silica.
The fifth is amber. Burmese amber has trapped insects, spiders, feathers, and even a small lizard from more than ninety million years ago, preserving the finest details of bristles and skin.
The 1993 film Jurassic Park persuaded many people that dinosaur DNA could be extracted from mosquitoes in amber. In reality, the apparently intact “soft tissues” in amber are often only cavities, films, or residual pigments. The DNA has not survived.
A perfectly preserved form does not imply perfectly preserved molecules. In 2015, Xing Lida of the China University of Geosciences spotted a piece of amber in the market at Myitkyina, northern Myanmar. It contained a small section of feathered tail. When the find was published in 2016, it was identified as the tail of a juvenile dinosaur, 99 million years old, with feather details and even the bones still inside.
The sixth route is freezing and drying. Siberian permafrost has yielded mammoths with fur and muscle intact; an Alpine glacier gave up the Iceman from 5,300 years ago.
The tissues may remain at a visible scale while their molecular information steadily degrades. In 2012, a study of 158 leg bones from extinct New Zealand moa calculated an average half-life of about 521 years for mitochondrial DNA fragments of a particular length. After each such interval, half the readable fragments are gone.
That is no universal figure for every environment. Low temperatures can extend it enormously. In 2022, environmental DNA was read from two-million-year-old sediments in northern Greenland, the oldest record currently known. Preserving a body and preserving readable genetic information are two different problems.
Beyond these routes, the environment itself does the choosing. At the La Brea Tar Pits in urban Los Angeles, animals have been getting trapped in asphalt for tens of thousands of years. The pits have yielded millions of fossils.
Ordinarily, herbivores should vastly outnumber carnivores. Here, however, about 90 percent of the mammal fossils are carnivores. One trapped bison would lure pack after pack of wolves and saber-toothed cats looking for a meal, only to trap them too. The list the pit has left us has been selected all over again by the trap.
One study compared living marine communities with a fossil database, asking whether each modern genus had any fossil record. About 15 percent did in open-ocean communities, about 34 percent in both shallow- and deep-water communities, about 44 percent on coral reefs, and about 51 percent on seamounts.
Shallow and deep water came out much the same. What often matters most is whether burial can keep pace with destruction. Nor is a lack of oxygen enough. It may drive away large scavengers, but anaerobic microbes keep decomposing the remains. Experiments show that oxygen deprivation alone cannot preserve soft tissue; early mineralization, adsorption onto clay, or rapid dehydration must also play a part.
Body size and abundance select too. An analysis of 9,408 marine animal genera over approximately the past 485 million years found that small-bodied genera were generally less likely to be sampled.
Abundant groups yield more specimens, and their first and last appearances in the record fall closer to their actual beginnings and ends. Rare groups appear late and vanish early. We will need this point again in Section V.
IV How Much Is Missing?
So how much has slipped through? Start with an old estimate. In 1977, taking the roughly 1.24 million living animal species then known as a baseline, a researcher estimated that only about a hundred thousand, some 8 percent, had a good chance of becoming fossils. That means “a good chance,” not “certain to become a fossil eventually.”
Behind this approach lies a list that one man spent more than twenty years compiling. Beginning in the 1970s, Jack Sepkoski in Chicago combed the paleontological literature, recording the first and last appearances of every marine animal genus on cards and in computers. Eventually he had assembled more than thirty thousand genera.
He died suddenly in 1999, aged fifty. His colleagues edited and published the list three years later. Today’s Paleobiology Database contains more than a million fossil occurrence records, entered one at a time by thousands of researchers around the world.
A more useful approach is to change the question: how likely is a group to be sampled within a given interval? In 1999, Michael Foote and Sepkoski at the University of Chicago devised a method, using more than sixteen thousand of roughly twenty-four thousand genera, with time intervals averaging about 5.3 million years.
It works like taking attendance. A student is present on Monday and Wednesday but absent from Tuesday’s register. He was probably there on Tuesday too; his name simply went uncalled. If a genus appears in the first and third intervals but not the second, it was plainly still alive in the second: it just was not sampled. Count how often this happens, how often something known to be alive goes unrecorded, and you can work backward to the sampling probability for each interval. You do not need to know how many genera originally existed in the world.
The differences were startling. Brachiopods had about a 0.9 probability of being sampled in each interval, cephalopods about 0.8 to 0.9, and trilobites about 0.7 to 0.9.
For polychaete worms, marine bristle worms such as ragworms, the figure was only about 0.05. On the same timescale, the best-recorded groups and one of the poorer ones differed by nearly twentyfold.
The brachiopods’ place at the head of the list has something to do with the calcite shells discussed in Section II. Their shells are hard and stable, and the animals often live in extensive beds on shallow seafloors. They look like clams, with two shells, but belong to a different animal phylum: their shells are upper and lower, rather than left and right.
They were everywhere in the shallow seas of the Paleozoic. Today only a few hundred species remain, tucked away in cold and deep waters. Going unrecorded in one interval out of ten is a very different predicament from being recorded in only one out of twenty.
Groups without hard shells, visible only through a few exceptional windows, often cannot support this calculation. The attendance method needs enough cases of “present in the first and third intervals, absent in the second.” When those patterns are too scarce, the problem is not merely a low sampling probability: this method cannot estimate it reliably. A quantity being small and the available data being insufficient to estimate it are different things.
Some groups fare far worse still. More than 90 percent of a jellyfish’s body is water. It dissolves within days of death, and only a handful of localities in the entire fossil record have yielded remains recognizable as jellyfish.
At Mazon Creek in Illinois, estuarine mud more than three hundred million years old holds thousands upon thousands of iron-rich nodules. Crack one open and you may find a flattened jellyfish. That is an exception, not the rule. Birds preserve poorly too: their bones are thin and hollow, easily crushed. Many bird fossils consist of just a few scattered bones.
There are other measures of completeness. Does a living family have at least one fossil member? That proportion is usually higher than the equivalent for genera or species. Nor does the presence of a genus in the record mean its skeleton is complete. “We have found this animal” and “how much of its anatomy we know” are separate matters.
Even the rocks themselves do not follow a simple rule of “the older, the fewer.” The University of Wisconsin’s Macrostrat database brings together about 33,900 stratigraphic units.
Plot the amount of rock against age and the curve has steps, peaks, and troughs, following the formation and disappearance of basins and the advance and retreat of seas. Even if the production of new organisms proceeded entirely smoothly, fluctuations in the amount of rock available for sampling would put peaks and troughs into fossil counts.
Whole stretches of pages are missing between rock layers as well. A famous boundary in the walls of the Grand Canyon places sandstone more than five hundred million years old directly on ancient rocks well over a billion years old. The interval between them, from hundreds of millions to more than a billion years, has left not a single page there.
As for “what fraction of all species that ever existed have we found as fossils,” there is no credible global figure. The total number of species that have ever lived depends on billions of years of speciation and extinction rates, on how species are defined, and on vast numbers of lineages never sampled at all.
Living species offer no impartial benchmark either. They are themselves survivors selected by the sieve of past extinctions. Familiar claims that “we have found only one in a thousand” or “only one percent” are calculations under particular assumptions. They are not measurements.
V How the Curve Gets Reshaped
The sieve does more than reduce the numbers. It reshapes the pattern.
The first effect has a name. In 1982, the American researchers Philip Signor and Jere Lipps set it out systematically, and it later took their names. Imagine a hundred species all disappearing on the same day.
Some have common fossils, others rare ones. The common ones can still be found in rocks laid down close to that day. The rare ones may last have been preserved long before: if you see something only once every ten years, you will not know that it died three years ago. Plot the curve and it shows “gradual decline.” In reality, everything disappeared in a single day.
One researcher experienced this effect firsthand. In the 1980s, Peter Ward of the University of Washington collected ammonites at Zumaia on Spain’s Basque coast.
The cliffs there lay out the final rock layers of the Cretaceous beside the sea. At first, he found ammonite species dropping out one after another tens of meters below the boundary. He concluded that the ammonites had been dwindling gradually.
Then he went back and searched more carefully, layer by layer. Right next to the boundary, he found species after species again. He changed his mind: the ammonites had probably disappeared together at the boundary. The earlier “decline” was down to his own insufficient sampling.
Dinosaurs have prompted the same argument. In Montana’s Hell Creek Formation, the last dinosaur bone lay about three meters below the boundary. Some took this as evidence that dinosaurs were already in decline before the great impact. In 1991, the American researcher Peter Sheehan led a large contingent of volunteers in an exhaustive search. They divided the rocks below the boundary into intervals and devoted equal effort to each. The results showed no gradual decrease in dinosaurs over the final two-million-plus years.
Start with a simplified model: occurrences enter the record independently at a constant average rate. The effect can be calculated. Suppose fossils of a group are sampled, on average, once per million years. The average gap between its actual disappearance and its last known fossil will then be about a million years; in 95 percent of cases it will be no more than about three million.
For a group sampled only once every ten million years, the average gap is ten million years. As we approach the actual endpoint, the number of known groups must shrink, because more and more lineages have already passed their “last time sampled.” “Fewer in each layer below the boundary” does not, by itself, demonstrate a gradual biological decline.
Turn the same logic around and you have delayed first appearances. A lineage has already originated; the first fossil that happens both to be preserved and to be found usually comes later.
The “oldest known fossil” sets a minimum age for an origin. Bats make a striking example. About fourteen hundred bat species live today, roughly a fifth of all mammals, yet their bones are delicate, many die in caves and forests, and their fossils are scarce.
The earliest known complete bat skeletons come from lake-bottom deposits in Wyoming about fifty-two million years old. At their first appearance, they are already flying bats, with wing skeletons much like those of bats today. The passage from flightlessness to flight is a blank in the fossil record.
The second effect is called the pull of the Recent. Younger rocks have suffered less destruction and are more readily exposed at the surface. More importantly, the very existence of a living group tells researchers that it has survived to the present.
Once a living genus has appeared even once in very old rocks, traditional counts treat it as present in every interval from that first appearance to today. Extinct genera receive no such free extension. The coelacanths we will meet in Section VII illustrate the point: more than sixty million years without a single fossil, yet their survival today pulls their stratigraphic range all the way to the present in one stroke. The closer the curve gets to modern times, the farther it is lifted.
How far has to be measured. In 2003, David Jablonski and colleagues at the University of Chicago examined living bivalves. Of 958 living genera and subgenera with fossil records, 906 had in fact already been sampled from rocks of the past five million years.
Remove the part “filled in solely because they are still alive,” and the pull of the Recent explains only about 5 percent of the rise in bivalve diversity during the Cenozoic. The bias is real. In this group, it cannot manufacture the entire trend.
The third effect is the amount of rock. In 1976, David Raup at the University of Rochester pointed out that marine fossil diversity through time correlates strongly with the quantity of sedimentary rock preserved.
One explanation is that more rock means more opportunities to sample, and therefore more fossils.
Another is rising sea level. During the Cretaceous, the sea once flooded the interior of North America, running from the Gulf of Mexico to the Arctic Ocean and splitting the continent in two.
Beneath the wheat fields of Kansas today lie mosasaurs and giant fish from that inland sea. Rising sea level expands the area of shallow seas, both allowing a real increase in marine life and leaving more marine strata. Both effects are real.
Simply dividing diversity by the amount of rock may remove the genuine ecological signal along with the bias. Comparisons of models show that the two relationships can coexist.
The fourth effect comes from exceptionally good fossil sites themselves. A compilation of about 694 exceptionally preserved fossil assemblages, ranging from 610 million to three million years ago, found that they cluster in both time and space.
A preservation window opens and large numbers of previously invisible soft-bodied forms enter the record together. It looks like a burst. The window closes and they all leave together. It looks like a decline.
The fifth effect comes from naming itself. In 1879, the American paleontologist Othniel Marsh named Brontosaurus. In 1903, another scholar judged it to be the same kind of animal as Apatosaurus, which Marsh had named two years before Brontosaurus. Under the rule of priority, Brontosaurus lost its name.
For more than a century, however, museum labels and postage stamps went on displaying it. In 2015, a detailed comparison of dozens of specimens concluded that the differences were great enough, and Brontosaurus was restored as a genus in its own right.
A 2002 study using North American fossil mammals estimated that about 24 to 31 percent of currently accepted species names might eventually prove invalid. Diversity could consequently be overestimated by more than 30 percent. The reverse happens when several real species are stuffed into one ill-defined name, depressing the count. Fossil species are distinguished by the forms that can be preserved; biological species also depend on the capacity to interbreed. Several species with identical hard parts can count as only one in the fossil record.
These effects alter different things. Premature last appearances and delayed first appearances shift endpoints. The pull of the Recent changes the slope. Rock quantity changes the rise and fall of the whole curve. Clusters of exceptional preservation change the intervals in which particular forms are visible. Naming changes the unit being counted. A corrective method usually addresses only one of these. So the questions are: which quantity has the bias changed, and by how much?
VI Can a Sieve Produce a Peak?
The first essay in this series already touched on a question: can a peak be entirely the work of the sieve?
First, set out the relationship. The observed number is approximately the product of four things: the true rate at which events occur, the probability of preservation, the probability of later exposure at the surface, and the probability of collection and identification. Every factor changes with age.
A useful rule follows. If the true occurrence rate is constant, preservation becomes less likely with increasing age, and the other two factors remain unchanged, the observed distribution can only decline from younger to older. Its highest point will be at the youngest end; it cannot produce a peak in the middle. A peak there requires something else: fluctuations in the true occurrence rate, rises and falls in preservation or exposure probabilities, several processes superimposed, a truncated time window, or ages that have been mixed or reset.
Essay I discussed the peak in lunar rock ages clustered around 3.9 billion years ago. Under this rule, preservation can produce it, for the last reason on the list: later impacts repeatedly shatter and remelt ancient impact melts, turning their isotopic clocks back again and again. This is the resetting of ages, not a simple case of “the older, the fewer.”
Something similar happens in shell accumulations on the seabed. Radiocarbon-date the shells one by one and you find that the longer old shells have lain there, the more have been ground up or dissolved. A reconstructed peak suggesting “especially abundant shellfish during this interval” is consequently pulled toward the present. What shifts is the timing of the peak itself.
Rock ages face the same problem. The oldest mineral grains on Earth are zircons about 4.4 billion years old, found in ancient conglomerates in the Jack Hills of Western Australia.
The hundreds of thousands of zircons dated around the world are not spread evenly across time. Their ages cluster in several peaks, often coinciding with the assembly of supercontinents. One school of thought takes those peaks to mark episodes when the crust was produced in great quantities.
Another argues that rocks formed during continental collisions are buried deeply and are therefore less easily destroyed: preservation made the peaks. What divides the two camps is precisely the question of this section.
We need examples pointing the other way as well. Rarefaction was not invented for fossils. In 1968, the American marine ecologist Howard Sanders wanted to compare worms and shellfish in deep- and shallow-water mud. Different places yielded different quantities of mud, so he devised a way to put every sample on the basis of the same number of individuals before comparing them.
In 1975, Raup formally applied the method to fossil diversity. Birdwatching makes the principle easy to grasp. One person watches in a park for ten hours, another for a hundred. The second is bound to record more bird species.
To compare two parks, you have to cut both observers back to the same effort, perhaps counting only their first two hundred birds. When Raup did the equivalent, the post-Paleozoic rise in the number of sea urchin families remained. That increase was real.
A better way to put samples on equal terms came out of codebreaking. During the Second World War, Alan Turing and his assistant I. J. Good at Bletchley Park devised an estimator to help decipher German codebooks. Count how many of the codes already intercepted occur only once, and you can estimate roughly what proportion remains unseen.
Good published the method after the war. He later advised Stanley Kubrick, helping him imagine the talking computer in 2001: A Space Odyssey. Paleontologists today use the same idea. Instead of requiring two intervals to yield the same number of specimens, they require the samples to “cover the same proportion of the underlying population.” It is fairer than fixing the number.
Several other methods have their own strengths. One uses stochastic processes to estimate origination, extinction, and sampling rates jointly. First and last appearances are no longer treated as actual beginnings and ends.
Another uses rocks without fossils as a control. If even the amount of unfossiliferous rock rises and falls with the diversity curve, the sampling explanation gains strength. A further approach asks whether independent kinds of evidence, with no shared sampling mechanism, turn at the same point: diversity, morphological disparity, ecological function. Conversely, if only one database’s counts show a peak, and that peak coincides exactly with a peak in sampling intensity, the observation process should be the first suspect.
All these methods share a boundary: a model can correct only for absences represented in its parameters. If you know roughly how often each group is sampled per million years, you can estimate endpoints that went unseen.
If you know the area of rock, you can include it in the model. But if a preservation process has left no recognizable trace of any kind, and there is no independent control, the data themselves cannot tell the model how many groups have remained entirely invisible.
Even unlimited data cannot determine that quantity, because those data contain no information about it. Statisticians call this non-identifiability. The global genus-level estimate discussed earlier excluded extremely soft-bodied groups; models that estimate rates jointly likewise begin with groups “recorded at least once.” Both methods stop in the same place. That is the boundary.
VII Four Kinds of Absence
We can now separate the meanings of “not there.” The four kinds that follow are this essay’s own arrangement of ideas scattered through the literature on stratigraphic statistics, taphonomy, and the philosophy of science.
The first: it never existed. The second: it existed, but left no remains that could be found. In the data, these can look exactly alike. Both have zero records.
Shortly before Christmas in 1938, Marjorie Courtenay-Latimer, a young curator at the East London Museum in South Africa, spotted a fish in a trawler’s catch. It was more than a meter long, its scales gleamed blue, and its fins looked like short, fleshy legs. She persuaded a taxi driver to take the huge, fishy-smelling thing back to the museum. But there was not a freezer in town big enough to hold it. Its internal organs could not be saved; she had to have it mounted. The chemistry professor J. L. B. Smith recognized it: a species of coelacanth.
Even the youngest known fossil members of this group were more than sixty million years old. People had long assumed that coelacanths died out at about the same time as the dinosaurs. They had been there all along, in the deep waters of the western Indian Ocean. They simply had left no fossils, and nobody had gone there to look.
In 1997, an American biologist and his new wife saw another species of coelacanth in a fish market on the Indonesian island of Sulawesi. Paleontologists call a group that disappears from the record and then returns a “Lazarus taxon,” after the biblical Lazarus, raised from the dead.
Plants have their examples too. In 1994, an Australian national park ranger found several tall conifers he had never seen before in a deep gorge northwest of Sydney.
They belonged to an ancient group of conifers common in the fossil record and thought to have vanished long ago. Only a few dozen mature trees survive in the wild, and their location remains secret. During Australia’s 2019–20 bushfire season, flames reached the edge of the gorge. Firefighters descended by helicopter, laid irrigation pipes, and watered the trees; aircraft dropped fire retardant. The trees were saved.
To distinguish these first two kinds of absence, we need something else: if the target existed, how confident should we have been of finding it? That can be calculated directly. Each independent survey has some probability of detecting a target that is present. The probability of missing it repeatedly is the product of the probabilities of missing it on each occasion.
If the detection probability in a single survey is 90 percent, then, assuming the target is present, the chance of missing it in all three surveys is just one in a thousand. Here, “not found” is strong evidence for “not present.” If the detection probability is only 1 percent, then, assuming the target is present, the chance of missing it in ten surveys is still about 90 percent. Both are cases of “repeated searches finding nothing,” but the probabilities of missing the target differ by nearly three orders of magnitude. The evidential weight differs enormously too. That is not a 90 percent probability of continued presence after the searches; that question also depends on the grounds for believing it was present before searching.
The evidential force of finding nothing depends on “how likely we should have been to find it if it existed.” The maxim “absence of evidence is not evidence of absence” therefore goes too far. Whenever evidence is more likely to be found if a hypothesis is true than if it is false, failing to find it weakens that hypothesis to some degree. The question is how much. Stratigraphic confidence intervals long ago turned this principle into a tool: when a large, evenly distributed run of fossils suddenly stops, the actual endpoint can be tightly bounded. The disappearance of a species that appeared only once every ten million years scarcely bounds anything at all.
The third kind: remains survive, but they do not carry the property we are asking about. Bones can answer questions about body size, joints, and growth. Usually, they cannot tell us the color of the skin.
What can be read falls roughly into three tiers. The form and dimensions of hard tissues, teeth, joints, skeletal proportions, bone histology, wear, injuries and disease, locality, and stratigraphic position can usually be read directly.
The structure of skin and feathers, some colors, muscles and internal organs, diet, locomotion, growth rate, and reproductive behavior depend on exceptional preservation or indirect proxies. Amber preserves the microstructure of the outer skin; early phosphatization preserves muscle; stable isotopes offer clues to physiology and environment.
Essay XVII will explain how the colors of dinosaur feathers can be read from melanosomes in fossils. Complete genomes, gene expression at a particular moment, the state of the nervous system, and calls or behaviors that left no physical trace generally cannot be recovered.
Proteins are much more durable than DNA. In 2019, researchers read sequences of enamel proteins from a Gigantopithecus molar about 1.9 million years old, excavated from a cave in Guangxi. From them, they established that the giant ape was closely related to orangutans. The five-hundred-and-some-year half-life in moa bones describes the lifespan of information about a property, far shorter than the lifespan of the form itself.
The fourth kind: within the scope in question, the quantity being asked about is undefined. Essay IV argued that extinction rates during the Great Oxidation Event lack a denominator. The microbial fossils of that time are little spheres; we cannot tell whether they represent one species or ten.
The limit of that statement lies in “the fossils of that time.” As soon as researchers define a unit they can track, whether a morphospecies or something else, they can define rates at which it appears and disappears. The quantity is undefined because there is no consistent procedure yet for its numerator, denominator, or unit of comparison. No geological age inherently forbids us to discuss extinction.
Can the first kind of absence be distinguished from the second? A blank alone cannot distinguish them, but detection probabilities can do so indirectly. Suppose “control groups” of similar body size and shell type are well preserved in rocks of the same age and kind, while the target group is consistently missing. The explanation “it was there but left absolutely nothing” becomes increasingly strained.
Conversely, if none of the comparable soft-bodied animals has been preserved, the target’s absence tells us almost nothing. The same “last appearance” can carry evidential weights differing by more than an order of magnitude in two different groups.
VIII The Shape of the Sieve Can Be Calibrated
At this point, it is easy to slide toward a conclusion: the fossil record cannot tell us anything. A test in 1992 answered that concern.
Mark Norell and Michael Novacek of the American Museum of Natural History both spent years digging fossils in the field. A few years later, their team uncovered a dinosaur in Mongolia’s Gobi Desert crouching over a nest, its forelimbs spread protectively around the eggs, in the same posture as a brooding bird today.
For the 1992 test, they assembled twenty-four groups of vertebrates and two quite independent sets of information. One consisted of evolutionary family trees inferred solely from the shapes of bones: which animals were more closely related, and which branches had split off first. The ages of the rock layers played no part in constructing them.
The other was the order of those animals’ first appearances in the strata. The two orderings were constructed separately: rock ages had not been fed into the trees inferred from bones. Both nevertheless reflect the same evolutionary history. If fossil first appearances were merely random noise, there would be no reason for them to agree with the branching order inferred from anatomy.
In eighteen of the twenty-four groups, the two orders agreed significantly. One measure used the shapes of bones, the other the sequence of rock layers. Two independent rulers measured the same order of events. In 2000, Michael Benton and colleagues at the University of Bristol expanded the test to a thousand evolutionary trees. They found that the agreement did not deteriorate appreciably in older rocks.
The record is broken, but not so broken that it has scrambled the large-scale sequence. The order established from fossils is broadly trustworthy. Caution is still needed when evolutionary trees are used in turn to “fill in” the fossil record: poor sampling can make the trees themselves unreliable.
Living and dead communities can also be compared. For years, Susan Kidwell and colleagues at the University of Chicago have counted living shellfish and the empty shells accumulated on coastal seafloors separately.
Even after decades to thousands of years of mixing, shell accumulations rank species by abundance in clear agreement with the living communities. In a synthesis of eighty-five datasets, more than 90 percent of the comparisons showed a positive correlation. Filtering had occurred, but it had not reduced the original ecological structure to a jumble.
The places where living communities and dead shells disagree carry information too. Kidwell found that along coasts altered by sewage, trawling, and aquaculture, shell accumulations often record the community that was there before people arrived. Dead shells on the seabed have become an old ledger nobody ever wrote. There is a great distance between “biased” and “devoid of information.”
In 1982, Raup and Sepkoski used Sepkoski’s list of marine families to chart the five great mass extinctions of the Phanerozoic. For more than forty years, that curve has been repeatedly interrogated for every bias discussed here, using rarefaction, coverage, rock quantity, and the pull of the Recent. Many details of the curve have changed. Not one of the five mass extinctions has disappeared.
A discipline follows from this: sampling bias cannot serve as a universal explanation. If any peak, trough, radiation, or decline can be attributed to sampling after the fact, then “sampling” explains nothing.
The better question is specific: for this group, on this timescale, in this environment, for this trait, and under this sampling design, which property of the real signal has the record preserved, and how large is the error? It is a cumbersome question, but an answerable one. Within the same global fossil record, some groups have about a 0.9 probability of being sampled in each interval, others about 0.05. “How complete is the fossil record overall?” has no single meaningful answer.
Return now to this essay’s construct. A construct, as this series uses the term, comes down to drawing a line: what is inside counts; what is outside does not. People may draw the line, as with a set of criteria, or nature may draw it, as with a membrane. Here the sieve draws it. Whatever passes through every opening enters the record and counts. Whatever fails to pass is, in that record, absent.
The previous ten essays encountered several kinds of absence. They are not the same. Essay Seven proposed downsampling the Phanerozoic record to make it more comparable with the Mesoproterozoic, then asking how much difference remains. Such methods can move the question forward, but the data must be prepared and the comparison’s assumptions made explicit. More than one explanation may survive. This gap can be narrowed through further work; we cannot promise to erase it in one exercise.
The dozens of origins of multicellularity in Essay VIII that never went far are fully visible; they were simply ranked farther down by a ruler with no natural unit. The Ediacaran organisms in Essay IX have thousands upon thousands of specimens, but every box in the grid demands a combination of traits they had not yet assembled. In the Cambrian diversity of Essay X, the contributions of real radiation, skeletonization, windows of soft-body preservation, and rock quantity cannot be disentangled cleanly.
What the construct cannot accommodate, yet cannot eliminate, this series calls the remainder. Here the remainder is the vast absent majority. In earlier essays, the remainder lay in coordinate axes, in boxes, in coefficients that could not be cleanly separated.
In this essay, the remainder lies in the sieve, and the shape of that sieve can be calibrated. Last appearances have confidence intervals. Genus-level sampling rates have the attendance method. Diversity has rarefaction and coverage. Origination, extinction, and preservation can be estimated together. The amount of rock can be measured from geological maps, and research bias can be measured against the area of exposed rock.
This sieve is an observation process that can be partly calibrated, not a black box. Every tool above was forced into existence by some misreading: Ward’s ammonites at Zumaia, Raup’s sea urchins, Jablonski’s bivalves. Each time the sieve has fooled us, we have measured a little more of its shape.
Keep hold of “partly.” We can calibrate what left some kind of trace. For things that left no trace through any channel, this record cannot give us the total. The denominator is not necessarily nonexistent; it is unknown. We therefore cannot calculate a percentage missed from this record alone. Something may have been there, or nothing may have been there. These traces will not tell us.
The criteria keep sifting. The correction models keep changing. But for what left no trace at all, the surviving record still supplies no total we can calculate.