Kew's plant report ran a flowering detector over 8 million herbarium sheets in a week

The Royal Botanic Gardens, Kew has published a State of the World's Plants and Fungi report that examines what digitization and machine learning are doing to botany, with contributions from 400 researchers in 40 countries ^[1][2]. The headline number inside it is an operational one: a classifier trained at the NTNU University Museum in Trondheim scored 8 million digitized herbarium specimens for whether the plant was in flower, a job its authors put at roughly 40,000 human hours ^[5][6][7].

That is about 20 working years of trained-eye labour, or a nominal 2,000 hours a year ^[8][11]. The machine took one week, according to postdoctoral researcher David Williamson ^[7]. Worth doing the division: 40,000 hours across 8 million sheets is 18 seconds per sheet ^[10]. The baseline is therefore a fast, narrow visual call on an already-photographed image, not curation, not identification, not label transcription. The saving is real, and it is a saving on one specific step.

The scale is also a fraction of what is sitting there. James Speed, a plant ecology professor at the same museum, says more than 145 million plant and fungi preparations have been digitized worldwide, across more than 170 institutions in 40 countries ^[3][4]. The 8 million sheets processed are about 5.5 percent of that ^[9], covering 200,000 species, an average of 40 sheets per species ^[6][12]. Williamson notes the technologies used are barely five years old, applied to collections built over centuries ^[24].

The result: global flowering times have shifted by an average of 2.5 days per decade over the past century, per Speed ^[13]. Over a century that is on the order of 25 days ^[23]. Flowering has moved both earlier and later, with the largest differences in the tropics ^[14]. Whether the 2.5 days is a net shift or an average magnitude of movement in both directions changes how you read it, and the material as supplied does not settle that ^[13][14]. The tropical finding surprised the team, because the largest warming has been in the far north ^[15]. Speed's explanation is that tropical flowering tracks precipitation, and a rainy season can arrive earlier, later, or not at all ^[16]. The consequence flagged in the report is a timing mismatch between flowering and pollinating insects, with knock-on effects for insect-eating birds and food production ^[17].

The same pipeline is aimed at the extinction accounting problem. More than 90 percent of fungal species remain unmapped ^[18], and only 1,000 plant species have been formally declared extinct, a figure the report treats as far too low ^[19]. Statistical models can estimate the probability that a species is extinct rather than merely undiscovered, and models can flag apparently unknown species in digitized collections for expert review ^[20].

Both quoted scientists put a ceiling on this. "Machines can't do the work of biologists, but they can help experts analyze datasets that would previously have taken a lifetime," Williamson said ^[21]. Kew taxonomist and report co-author Martin Cheek was blunter: "I think the potential for AI is enormous, but it is still currently potential" ^[22].

What to watch: whether the flowering classifier's error rate and validation set are published, since a phenology signal of 2.5 days per decade sits close to the plausible noise floor of a binary image call ^[13]; whether the remaining roughly 137 million digitized sheets get processed or stall on institutional access terms ^[4][9]; and whether any of the model-flagged candidate species reach a formal description, which still requires taxonomists ^[20][22].