Florida Museum Study Proves AI Requires Far Fewer Fossils Than Expected to Revolutionize Species Identification

Florida Museum study finds AI only needs 250 fossils to accurately identify species. Discover how computer vision is clearing the backlog in paleontology.

By: AXL Media

Published: Feb 26, 2026, 6:31 AM EST

Source: The information in this article was sourced from Florida Museum of Natural History

Overcoming the Backbone of Paleontology's Data Crisis

Vertebrate paleontology has long grappled with a "numbers problem" that slows the pace of scientific discovery. Unlike fossilized pollen or spores, which are found in the thousands, vertebrate remains are typically discovered as isolated, broken fragments. A single skeleton consists of over 200 bones, yet finding a nearly complete specimen is a rarity. This fragmentation creates a massive backlog in museums; the Florida Museum alone houses over one million specimens, including thousands of bags of unsorted sediment. Manual identification of these fragments is the most time-consuming aspect of the field, creating a bottleneck that AI computer vision is now poised to clear.

The 250-Specimen Breakthrough

A new study co-authored by Distinguished Professor Emeritus Bruce MacFadden reveals that the amount of data needed to train a high-performing AI is much smaller than previously assumed. By testing shark teeth—which are durable and plentiful compared to other vertebrate bones—the team established a threshold for accuracy. They found that AI performance plateaus at approximately 250 specimens. Beyond this number, additional data yields only marginal improvements in accuracy. This finding is a game-changer for paleontologists working with rare species, as it proves that AI can be a viable tool even when fossil records are patchy.

High Accuracy with Minimal Training Data

The research team focused on six shark species from the Neogene period, including the extinct Megalodon and the modern Great White. Using computer vision models fine-tuned by software experts, the team tested the algorithms in increments of 50 images. Surprisingly, even at the lowest training level of just 50 specimens, the models achieved accuracy rates of at least 93%. This suggests that the "breakdown point" for AI in paleontology is much lower than expected, allowing for the rapid classification of fossils that were previously deemed too sparse for machine learning applications.