Kew Gardens scientists have used artificial intelligence to assess extinction risk across every known flowering plant species — a first — and the results confirm what many botanists had feared: Kew's State of the World's Plants and Fungi Report estimates 45% of flowering plant species are at risk.

The scale of that figure deserves a moment. Angiosperms — the flowering plants — comprise roughly 90% of all plant life on Earth. They underpin virtually every terrestrial food web, anchor most agricultural systems, and supply the majority of medicinal compounds in clinical use. A near-majority under threat is not a footnote. It is a structural problem for ecosystem function at the planetary scale.

What the AI methodology changes is the coverage. Conventional IUCN Red List assessments require taxonomic expertise, field data, and years of review per species — a bottleneck that has left the vast majority of plant species formally unevaluated. By training models on the assessed fraction and projecting across the full angiosperm tree, Kew scientists have produced the first comprehensive risk profile for the group, published in March 2024. The output is probabilistic, not a species-by-species verdict, but at this scale probabilistic coverage is precisely what conservation triage requires.

The gaps in baseline knowledge compound the risk picture. In 2025, Kew recorded 125 plants and 65 fungi named new to science — a reminder that the inventory itself is incomplete. More pointedly, a 2023 Kew analysis estimated that three in four undescribed plant species globally are already likely at risk before they have even been formally named. That figure inverts the usual discovery narrative: finding a new species no longer implies it is safe, or even abundant.

Geography structures the ignorance as much as the risk. Kew has identified 32 global plant diversity 'darkspots' — regions where species richness is high but taxonomic and ecological data are thin. These are the zones where the AI model's uncertainty is widest and where targeted survey work would yield the greatest reduction in assessment error. Many overlap with tropical and subtropical areas under active land-use pressure, which means the knowledge deficit and the threat are spatially correlated in the worst possible way.

The practical question for conservation planners is what the AI output is and is not useful for. It is not a substitute for site-level threat mapping or for the legal and policy frameworks that flow from formal IUCN listings. But as a prioritization layer — directing limited survey and assessment resources toward genera and regions where predicted risk is elevated — it fills a gap that has existed since modern conservation biology began. The 45% headline should be read as an upper-bound estimate under current trajectories, not a confirmed count; the confidence intervals matter and should be interrogated in the underlying methodology.

For practitioners working on the Convention on Biological Diversity's Kunming-Montreal targets — specifically Target 4, which calls for halting human-induced extinction of known threatened species by 2030 — the Kew AI work offers a new data layer for national biodiversity strategies. The harder problem is that Target 4's framing of "known threatened species" is precisely where the undescribed and unevaluated fraction falls through the floor. If 75% of undescribed plants are already at risk, and discovery pipelines are adding only a few hundred species per year to formal taxonomic records, the policy architecture is chasing a moving baseline.

That is the structural tension this research surfaces: conservation frameworks are built on inventories, and the inventory is demonstrably incomplete at the same time that the assessed fraction of it is deteriorating. The AI approach does not resolve that tension, but it makes the shape of the problem legible in a way it was not before.