Norman MacLeod and colleagues' call to develop automated species-identification systems is laudable (Nature 467, 154–155; 2010), but let's not forget a core feature of taxonomic work that depends on a scholar's input — the discovery of new characters.

Unexpected evolutionary novelties in morphology and physiology, for example, are what make taxonomic exploration rewarding. Taxonomists set out to discover and track such novelties and their evolutionary history. It is in this sense that taxonomy provides the empirical basis for understanding speciation and phylogeny.

There is a place for automated pattern detection, but it would not work with the 5,000 species of Drosophila, say, which are identified by their many different structures. Taxonomy can independently test and verify identifications without relying on patterns of single characters, as the long list of synonymies in any biological group testifies. The practice of taxonomic revision and publishing detailed monographs ensures that character distributions, species status and phylogenetic relationships are subject to repeated and critical testing.

We should beware the trend to confuse automatic identification tools with those that are useful for discovering new species. The emerging field of cybertaxonomy is an advance only if it is understood as enhancing and enabling theory-rich descriptive taxonomy, not replacing it.

As in many other modern scientific fields, including diagnostic medicine and molecular genetics, a final step involving a human expert is essential.