The fossil collection behind AI’s disputed history of sharks

A fossil inventory published on 8 October brings together 30,047 occurrences of sharks, rays and their relatives from 5,867 collections and 2,055 references. Called FINS, it covers 145 million years and records 1,621 species. Its contribution is a traceable account of where and when those animals have been found.[1]

That is valuable groundwork for AI because the history hidden in fossils is different from the history of collecting them. The inventory gathers records scattered through decades of literature, expanding geographical coverage, particularly in the Global South. Sharks and rays leave abundant teeth and denticles even though their largely cartilaginous skeletons preserve poorly, as participating university UCSC explained on 14 October.[2]

An occurrence records a taxon at a particular place and time, not individual animal abundance. FINS links occurrences, collections and references in three data files. It began with 2020 snapshots from the Paleobiology Database and Shark-References, followed by literature extraction and curation.[1]

Learning what an incomplete record can hide

The AI connection is DeepDive, a method introduced in 2024. It generates simulated histories of species diversity and then simulates the fossil records those histories could leave. The training worlds vary geographically and through time, including differences in preservation and sampling. A recurrent neural network learns to work backwards from the incomplete observations towards the underlying diversity trajectory.[3]

Imagine two intervals with the same number of recorded species. One might contain plentiful, widely sampled fossil sites; the other might be represented by a few accessible outcrops. Equal observed counts would not necessarily imply equal original diversity. Simulation provides examples in which both the real population of species and the sampled record are known, allowing that distinction to become a learning problem.[3]

In the January Current Biology study that used an earlier version of FINS, DeepDive reconstructed shark and ray diversity reaching a peak during the Lutetian stage of the Eocene, about 47.8–41.2 million years ago. The estimated net decline from that peak to the present was 41.3%. The same analysis estimated a much smaller 9.7% net fall across the intervals surrounding the end-Cretaceous extinction.[4]

The latter number spans the Maastrichtian and Danian, approximately 72.1–61.6 million years ago. It measures a change in reconstructed species richness across those intervals. It does not mean that only 9.7% of species went extinct at the moment of the asteroid impact, or that a measured fraction of individual animals survived.[4]

A simple illustration shows why. If a group starts with 100 species, loses 30 and gains 20, it finishes with 90. The net loss is 10%, while 30% of the starting species disappeared. Origination and extinction can partly offset each other in a richness curve.

The teeth have names, and the names matter

The reconstruction has prompted a substantive scientific dispute. In an April preprint, Guillaume Guinot and colleagues argued that taxonomic assignments, fossil ages and the treatment of uncertain identifications affected the result. A name accompanied by “aff.” or a question mark can carry uncertainty that disappears when a database treats it as an unqualified species record. Such decisions can change whether a species appears to cross an extinction boundary.[5]

The original authors responded with data checks and alternative analyses. Their 10 April preprint reported larger net declines under more restrictive treatments:[6]

The latest response abstract, dated 24 July, acknowledges valid corrections alongside disagreements over data treatment and proposed changes the authors could not verify. They maintain that the tested alternatives do not support catastrophic net diversity loss. Both the critique and response are preprints; the numerical comparison above is explicitly the April version.[7]

An inventory that can be questioned

FINS supplies a practical way to investigate such choices. Its app supports filtering, alternative taxonomic treatments and export of customised datasets; its curation records let users trace decisions back to source material.[1]

UCSC’s account also stresses that specialists must critically validate new entries, even where software assists with adding geographical, temporal and taxonomic information. The collection can grow as additional records enter the literature and databases.[2]

For an AI reconstruction, this creates a useful division of work. A model can compare an incomplete record with many simulated histories. A palaeontologist can examine whether a tooth belongs to the named species and whether its assigned age is sound. A shared, traceable inventory lets those two forms of inquiry meet on the same evidence.

Report an error or suggest a correction