Machine learning has helped researchers trace schooling behaviour across 2,839 fish species, finding that it evolved independently about 60 times and was lost around 320 times. The study’s bigger lesson is that behaviour once considered too slippery to compare at scale can be investigated across the evolutionary tree.
Watch Desk analysis
What happened
The research, published in the Proceedings of the National Academy of Sciences on 6 November, combines machine-learning analysis of written species descriptions with computer analysis of fish images and videos. The team used these methods to estimate schooling behaviour in nearly 10% of the 39,000 living fish species.
Across roughly 350 million years of fish evolution, the researchers found repeated gains and losses of collective movement. Species lower in the food chain were more likely to school, though the relationship was far from universal. The researchers also found that lineages whose species all school were more diverse, but could not tell whether that reflected faster speciation or lower extinction. The Cornell Chronicle’s report describes the study and its methods.
Key findings
- Schooling evolved repeatedly
The researchers estimate about 60 independent gains and 320 losses across the fish evolutionary tree. - Machine learning widened the sample
Text analysis and image-based movement estimates helped assess schooling in 2,839 species, far beyond the handful typically examined in earlier studies. - Predator pressure is part of the picture
Fish lower in the food chain were more likely to school, but many exceptions point to other influences. - Schooling lineages were more diverse
The study found greater diversity where all species in a lineage school, but could not separate higher speciation from lower extinction.
Why it matters
Behaviour is harder to compare across species than a physical feature you can measure in a specimen. The researchers’ approach offers a way to turn scattered descriptions and images into a much broader dataset, giving evolutionary questions a scale they have rarely had.
It also complicates the tidy story that schooling is simply a defence against predators. The association is there, but schooling has been gained and lost repeatedly, and the study points to possible trade-offs including competition, misinformation and disease risk.
Our read
This is a strong example of machine learning doing useful scientific work: helping researchers ask a question across thousands of species, rather than merely automating a familiar task. The findings are correlations, not a complete explanation of why fish school, but the scope of the analysis makes the next questions much better grounded.
What to watch
- Whether researchers apply similar methods to other animal behaviours at comparable scale.
- How future work tests the proposed costs of schooling, including competition and disease.
- Whether the link between schooling lineages and diversity holds when speciation and extinction are measured separately.
Discussion spark: If machine learning can make behaviour measurable across thousands of species, should researchers trust those large-scale patterns before the underlying observations can be checked species by species?
Sources and evidence
- Source update (9 November 2026, 14:27 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.