The Dinosaurs Were Fine. The Rocks Weren't.

Something Everyone Half-Already Believes

Ask a room of people who like dinosaurs one question — were they already in trouble when the asteroid hit, or was it a clean kill on a thriving group — and you will get a confident answer either way, usually delivered as settled fact. That confidence is the interesting part. The underlying data are genuinely contested, have been for over forty years, and the argument keeps getting re-litigated with new statistical tools rather than new fossils.

Nobody disputes the ending. A bolide struck what is now the Yucatán Peninsula 66 million years ago and non-avian dinosaurs do not appear in the fossil record afterward, anywhere. The disputed part is the several million years immediately before that, specifically in the Campanian and Maastrichtian of western North America, where the best terrestrial fossil record of the whole Mesozoic happens to sit. Plot raw genus or family counts against time in that record and, over and over, you get a rise into the mid-Campanian, a peak somewhere around 75 million years ago, and a decline running into the boundary. Whether that decline is telling you something true about the animals, or something true about the rocks they are buried in, is the entire question this field note is about.

Counting a Species That Has Been Dead for 66 Million Years

Start with the basic problem, because it is easy to underrate. You cannot survey Maastrichtian dinosaurs the way you would survey birds in a forest. You cannot go back and look. What you have instead is a set of occurrences: a bone, or a bonebed, or a footprint, found at a mapped point, in a rock unit that can be dated (usually by its position relative to volcanic ash beds that can be radiometrically dated, or by magnetostratigraphy, or by correlating fossil pollen and other biostratigraphic markers between sites), assigned to a taxon by whoever described it.

A "diversity curve" is built by binning those occurrences into time intervals and counting how many distinct taxa show up in each bin. That sounds straightforward until you notice everything the count depends on besides biology: how much rock of the right age is exposed at the surface anywhere in the world today; how much of that rock has ever been walked by someone looking for bone; how good the preservational conditions were at the time of burial; how finely taxonomists have chosen to split or lump the specimens found; and how much fieldwork funding existed in a given decade for a given formation. None of that is dinosaur biology. All of it moves the count.

This is not a new worry. Philip Signor and Jere Lipps gave the general version of it a name in 1982: even if a group actually vanished suddenly, the observed last occurrences of its member species will be scattered backward in time from the true extinction point, purely because sampling is imperfect and each species' true last-known fossil is a random draw before its real disappearance. The result, plotted naively, looks exactly like a gradual decline leading into a sudden event, whether or not one occurred [2]. The Signor–Lipps effect is the reason nobody in this field gets to say "the counts went down, therefore the animals went down" and expect to be taken seriously.

Same Family, Same Facies, No Decline

The first serious quantitative attack on this problem, and the direct intellectual ancestor of the 2025 paper, is nearly as old as this reporter's parents. In 1991, Peter Sheehan, David Fastovsky and colleagues published a three-year field study of dinosaur family-level diversity through the Hell Creek Formation of Montana and North Dakota in Science [3]. Their trick was to control for the thing everyone suspected was confounding the counts: not every part of a formation preserves fossils equally well, so comparing raw counts across a formation's full thickness mixes together facies (rock types formed under different conditions) with genuinely different preservation odds.

So they split the Hell Creek Formation into three roughly equal stratigraphic intervals and, within each interval, compared only similar facies to similar facies — channel sandstone to channel sandstone, floodplain mudstone to floodplain mudstone — rather than lumping everything together by elevation in the section. Family-level diversity, measured this way, showed no statistically significant decline through the run-up to the boundary (p < 0.05) [3]. Their conclusion was that the dinosaurs disappeared abruptly, consistent with the newly proposed impact hypothesis, not gradually. Dale Russell had already floated the same suspicion two years earlier in a short, skeptical piece in Nature asking outright whether the "gradual decline" was fact or fallacy, pointing at the geological record itself as the more likely source of the pattern [4].

Notice what the 1991 method actually is: a facies-matched comparison, done by hand, on one formation. It is careful and it holds up, but it does not scale to a continent, to multiple clades, or to a rock-exposure variable that changes continuously in space rather than sorting neatly into "similar" and "different" bins. Fastovsky and Sheehan revisited the same ground for a broader geoscience audience in a 2005 GSA Today feature, restating the case for an abrupt extinction rather than a long slide and summarising fourteen years of follow-up work on the question [11]. That limitation is exactly what the newer methods, including the 2025 paper, exist to remove.

Rarefy the Counts, Watch the Decline Move

Through the 2000s and 2010s the standard answer to sampling worries was subsampling: instead of comparing raw counts between intervals, draw equal-sized random samples of occurrences (or of formations, or of collections) from each interval, over and over, and compare the resampled richness. If a decline survives being rarefied down to equal sampling effort, the argument goes, it is probably not purely an artifact of unequal effort.

The most cited application of this logic to dinosaurs is Graeme Lloyd and colleagues' 2008 Proceedings of the Royal Society B paper on what they called the Cretaceous Terrestrial Revolution [5]. Their headline result was not about the final decline at all — it was that the celebrated mid-Cretaceous explosion in dinosaur diversity, the one usually credited to the spread of flowering plants, is itself largely a sampling artifact once you subsample and model diversification properly; branching rates look statistically constant across most of dinosaur history. As part of the same analysis, they found no evidence of a progressive decline in the run-up to the extinction either [5]. A separate methodological line, Jostein Starrfelt and Lee Hsiang Liow's 2016 TRiPS model (True Richness estimated using a Poisson Sampling model), pushed the same logic further by fitting an explicit statistical model of sampling intensity per taxon across the whole Mesozoic. It estimated roughly 1,936 dinosaur species overall and, more to the point, showed that every major dinosaur subgroup carries "considerable" sampling bias throughout its history, bias large enough to call even the timing of well-known diversity swings into question [6].

So by the late 2010s the technical literature had already mostly turned against a real terminal decline, using two independent correction strategies. Stephen Brusatte and ten co-authors said as much plainly in a 2015 Biological Reviews synthesis of the whole debate: after reviewing two decades of accumulated evidence, they concluded that "little support exists for a global, long-term decline across non-avian dinosaur diversity prior to their extinction," while still flagging a real, narrower signal — a reduction in the number of large-herbivore species specifically in North America in the last few million years [7]. That is an important nuance this field note keeps coming back to: "no global decline" and "no regional change at all" are not the same claim, and the 2025 paper's authors were about to sharpen exactly that distinction.

Verdicts condensed to one line each from the abstracts and results of the cited papers; see References for the full citations. This is the club's summary table, not a table from any of the papers.

The Nearest Relative of This Paper

One entry in that table deserves its own section, because it is the most direct scientific ancestor of the 2025 result and shares two of its authors. In 2019, Alfio Alessandro Chiarenza, Philip Mannion and colleagues took a completely different route into the same question: instead of modelling the fossils, they modelled the climate. Using a fully coupled climate model to reconstruct latest Cretaceous North American environments and an ecological niche model to project suitable dinosaur habitat onto that reconstruction, they found that continent-wide habitability was stable, or even increasing, from the Campanian into the Maastrichtian — a rise "that is not preserved" in the fossil record as we have it [8]. Their explanation for that mismatch was that the apparent decline reflects sampling bias driven by geological changes over the interval, specifically mountain building and sea-level shift altering how much of the right-aged rock is exposed and where. They concluded outright that Maastrichtian dinosaur diversity was likely being underestimated [8].

That is a striking result on its own, but it has an obvious gap: it tells you the animals probably had somewhere to live, without directly touching the actual mechanism by which the fossil record fails to show it. The 2025 paper is best read as filling exactly that gap — taking the "it's the rock, not the animals" hypothesis and testing it directly against the geological variable it blames, using the fossils themselves.

Borrowed From People Who Count Live Animals

Here is the actual trick, and it did not originate in paleontology. Ecologists who survey living species face a version of the exact same confound: if you walk a transect and fail to see a bird, that could mean the bird was not there, or it could mean the bird was there and you missed it. Confusing those two possibilities wrecks any comparison between sites or between years. Darryl MacKenzie and colleagues formalised the fix in a widely used 2002 Ecology paper: survey each site more than once, and use the pattern of detections and non-detections across repeat visits to estimate two separate probabilities at once — the probability ψ that a site is truly occupied, and the probability p that the species is detected given that it is present [9]. A single non-detection then tells you almost nothing on its own; the model only concludes "probably absent" once repeated failures to detect, at a site's own detection rate, become statistically unlikely under "probably present."

The basic relationship is simple to write down and does a lot of work:

Read it plainly: what you actually observe is the product of whether the animal was there and whether you could have found it there. A raw occurrence count only ever measures the left-hand side. If p falls over time for reasons that have nothing to do with ψ, the observed count falls right along with it, and nothing in the raw count alone can tell you which factor did the falling. You need an independent estimate of what should be driving p, fed in as its own variable, to pull the two apart.

What Dean and Colleagues Actually Built

That is the model the 2025 team imported into the Late Cretaceous. Christopher Dean, Alfio Alessandro Chiarenza, Jeffrey Doser, Alexander Farnsworth, Lewis Jones, Sinéad Lyster, Charlotte Outhwaite, Paul Valdes, Richard Butler and Philip Mannion built a Bayesian occupancy model over a grid laid across North America, tracking four major, well-sampled dinosaur clades — Ankylosauridae, Ceratopsidae, Hadrosauridae and Tyrannosauridae — through four time intervals spanning the last roughly 18 million years of the Cretaceous, from about 84 to 66 million years ago [1]. For each clade, in each grid cell, in each interval, the model estimates ψ: the probability the clade actually occupied that cell at that time. Separately, it estimates p: the probability of detecting that occupancy in the rock record we have access to today, and it lets p depend on covariates that have nothing to do with dinosaur biology — paleoclimate and geography reconstructed for the interval, the area of geological outcrop of the right age actually exposed at the surface, and modern land cover, which governs whether exposed rock is even visible today under present-day vegetation, ice, or development [1].

Two Curves That Should Agree, and Do Not

Run that model and two very different pictures come out for the same fossils. The raw occurrence record, treated the naive way every diversity curve has always been built, does what it has always done: it declines toward the boundary. The modelled occupancy probability ψ, the thing the raw count was supposed to be a proxy for, does not. Across the four clades, occupancy holds "generally stable" through the interval [1]. What tracks the decline almost exactly is the detection probability p — and the paper reports that this happens "despite high overall sampling during this interval," which is the detail that makes this something more than a restatement of "we didn't look hard enough." Effort was not obviously lacking. The rock available to look in was going away [1].

In the lead author's own words, describing the result to the press: "The probability of finding dinosaur fossils decreases, while the likelihood of dinosaurs having lived in these areas at the time is stable" [10]. That sentence is the whole paper compressed to twenty-two words.

Why the Rock Itself Got Worse

It helps to know, separately from any model, that rock exposure genuinely does shrink toward the top of a stratigraphic sequence for boring, non-biological reasons. Younger sediment has had less time to be uplifted and exhumed by erosion into something you can currently stand on. Basins that were actively subsiding and accumulating sediment in the Campanian were not necessarily still doing so, in the same place, by the Maastrichtian, as regional tectonics and sea level shifted the geography of where sediment was being laid down at all. Chiarenza and colleagues pointed at exactly this — mountain building and sea-level change reshaping where late Maastrichtian rock exists to be found — as the geological mechanism behind their climate-model mismatch [8]. The 2025 paper's contribution is to add a second, independent filter on top of the geological one: even where late Cretaceous rock is genuinely exposed at the surface today, whether a modern field crew can walk up to it depends on modern land cover — whether that outcrop sits under grassland easy to prospect, under forest, under a city, or under the ice and permafrost of higher latitudes [1]. Two different eras, 66 million years apart, both conspiring to hide the same rock.

The Ceratopsian Wrinkle

A model built entirely to explain away declines would not be worth much if it explained away everything, including real change. It does not. Reporting on the study specifically notes that ceratopsians — the horned, frilled clade that includes Triceratops — show increased detection later in the interval, and that this rise correlates with an environmental shift toward the open, plain-like habitats ceratopsians are thought to have favoured [10]. That is not the model finding "no change anywhere." It is the model finding a clade-specific signal that runs in the opposite direction from the shrinking-rock story and lines it up with an independent environmental cause, which is close to the best evidence you can ask an artifact-hunting method to produce: it does not just flatten every curve to zero, it can also tell you when something really did move.

Club Arithmetic: The Peak Sits Exactly in the Middle

Nothing in this section is the paper's. It is ours, done on numbers the paper and the coverage of it report, and flagged as such.

The study window runs from about 84 Ma to 66 Ma — an 18-million-year span [1]. The reported raw-occurrence peak sits at about 75 Ma [10]. Subtract: 84 minus 75 is 9; 75 minus 66 is also 9. The peak that anchors the entire "decline" narrative sits, to the precision these round numbers allow, exactly at the midpoint of the study's own window. Every account of "the decline" is therefore, by construction, a story about the second half of an 18-million-year interval compared against the first half — which is worth remembering the next time a headline treats 9 million years of falling counts as if it were the whole Cretaceous unravelling.

Coverage of the study also reports a dataset of more than 8,000 fossil occurrences analysed across that window [10]. Spread evenly across 18 million years, purely as an average and not a claim about how occurrences are actually distributed in time, that is a little over 444 fossils per million years — a single, rough yardstick for how much raw material an occupancy model like this one has to work with per time slice, and a reminder that "high overall sampling" in the abstract's own words is not a small dataset thinning out at the edges [1].

Already Declining, or Cut Off in Their Prime?

Here is why any of this is worth an argument that has now run for over forty years. If the pre-impact decline were real and global, the asteroid would be less a cause of extinction than a coup de grâce delivered to a group already losing ground, perhaps to competition, climate cooling, or the Deccan Traps volcanism erupting through roughly the same interval. If the decline is mostly a record artifact, the more striking and, to this club, more interesting story holds instead: a diverse, ecologically stable set of dinosaur clades, doing fine by any measure we can actually trust, cut off in a single geological instant by a rock arriving from space. Those are different stories about contingency in the history of life, and they carry different lessons about how vulnerable a "healthy" group can be to a large enough external shock.

The accumulating weight of the last four decades of methodologically serious work — facies-matched counts in 1991, subsampling and diversification modelling in 2008, an explicit sampling model in 2016, climate-and-habitat modelling in 2019, and now occupancy-versus-detection modelling in 2025 — leans toward the second story for North America specifically, with Brusatte and colleagues' 2015 caveat about real, narrower faunal turnover among large herbivores still standing [7]. What has kept the first story alive in the public imagination for so long is simply that the easiest thing to compute from a fossil database, a raw count by time bin, keeps producing the decline shape no matter how many times more careful methods argue it away.

What Would Still Change Our Mind

A few honest weak points, in the spirit of steelmanning rather than dismissing the older view. First, the study covers four clades and one continent; a reader should resist generalising "the dinosaurs were fine" to every lineage and every landmass on the strength of Ankylosauridae, Ceratopsidae, Hadrosauridae and Tyrannosauridae in North America alone. Second, an occupancy model is only as good as its detection covariates; if geological outcrop area and modern land cover do not fully capture what actually governs detectability (fossil-preparation backlogs, museum curation practices, and which formations happen to be fashionable to work on in a given decade are all plausible omitted variables), some genuine biological signal could still be riding along inside what the model currently assigns to "detection." Third, Brusatte and colleagues' regional herbivore-turnover finding [7] has not been directly re-examined against this occupancy framework in what we could access, so "no global decline" and "no change in large-herbivore composition" are not obviously the same claim, and this paper does not fully reconcile them.

What We Take From This

The part of this paper that stuck with us is not really about dinosaurs. It is that an entire scientific argument, run for four decades by capable people using increasingly sophisticated statistics, kept getting fooled by the same variable because nobody had a clean way to separate "the animal wasn't there" from "we couldn't have seen it if it were." That is not a paleontology problem specifically. It is the same problem camera-trap ecologists solved for live tigers, and the same problem epidemiologists face when case counts fall because testing capacity fell. The occupancy-versus-detection split is not a clever trick invented for dinosaurs; it is a general-purpose fix for exactly this confusion, arriving in paleobiology about twenty-three years after ecologists worked it out for themselves. We would like to know what else in the fossil record has been quietly tracking a sampling variable this whole time, waiting for someone to think to ask.

References

1. 1.

   Dean, C. D., Chiarenza, A. A., Doser, J. W., Farnsworth, A., Jones, L. A., Lyster, S. J., Outhwaite, C. L., Valdes, P. J., Butler, R. J. & Mannion, P. D. (2025). The structure of the end-Cretaceous dinosaur fossil record in North America. Current Biology 35(9), 1973–1988.e6. doi:10.1016/j.cub.2025.03.025

2. 2.

   Signor, P. W. & Lipps, J. H. (1982). Sampling bias, gradual extinction patterns and catastrophes in the fossil record. Geological Society of America Special Paper 190, 291–296. doi:10.1130/SPE190-p291

3. 3.

   Sheehan, P. M., Fastovsky, D. E., Hoffmann, R. G., Berghaus, C. B. & Gabriel, D. L. (1991). Sudden extinction of the dinosaurs: latest Cretaceous, upper Great Plains, U.S.A. Science 254, 835–839. doi:10.1126/science.11536489

4. 4.

   Russell, D. A. (1984). The gradual decline of the dinosaurs — fact or fallacy? Nature 307, 360–361. doi:10.1038/307360a0

5. 5.

   Lloyd, G. T., Davis, K. E., Pisani, D., Tarver, J. E., Ruta, M., Sakamoto, M., Hone, D. W. E., Jennings, R. & Benton, M. J. (2008). Dinosaurs and the Cretaceous Terrestrial Revolution. Proceedings of the Royal Society B 275, 2483–2490. doi:10.1098/rspb.2008.0715

6. 6.

   Starrfelt, J. & Liow, L. H. (2016). How many dinosaur species were there? Fossil bias and true richness estimated using a Poisson sampling model (TRiPS). Philosophical Transactions of the Royal Society B 371, 20150219. doi:10.1098/rstb.2015.0219

7. 7.

   Brusatte, S. L., Butler, R. J., Barrett, P. M., Carrano, M. T., Evans, D. C., Lloyd, G. T., Mannion, P. D., Norell, M. A., Peppe, D. J., Upchurch, P. & Williamson, T. E. (2015). The extinction of the dinosaurs. Biological Reviews 90, 628–642. doi:10.1111/brv.12128

8. 8.

   Chiarenza, A. A., Mannion, P. D., Lunt, D. J., Farnsworth, A., Jones, L. A., Kelland, S.-J. & Allison, P. A. (2019). Ecological niche modelling does not support climatically-driven dinosaur diversity decline before the Cretaceous/Paleogene mass extinction. Nature Communications 10, 1091. doi:10.1038/s41467-019-08997-2

9. 9.

   MacKenzie, D. I., Nichols, J. D., Lachman, G. B., Droege, S., Royle, J. A. & Langtimm, C. A. (2002). Estimating site occupancy rates when detection probabilities are less than one. Ecology 83, 2248–2255. doi:10.1890/0012-9658(2002)083[2248:ESORWD]2.0.CO;2

10. 10.

    Phys.org (2025). Dinosaurs' apparent decline prior to asteroid may be due to poor fossil record, say researchers. News report on [1], including comment from C. Dean.

11. 11.

    Fastovsky, D. E. & Sheehan, P. M. (2005). The extinction of the dinosaurs in North America. GSA Today 15(3), 4. doi:10.1130/1052-5173(2005)15<4:teotdi>2.0.co;2