Western AI fails on African farms, local models take root
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Catherine Nakalembe of NASA Harvest and the University of Maryland fitted GoPro cameras on dozens of helmet-wearing volunteers in western Kenya, collecting over 5 million images in two weeks to train AI to recognize maize, beans, and cassava — crops Western models couldn't identify
- Farmers for Forests in India's Maharashtra state found a popular open-source AI 'failed spectacularly' at tree detection, missing over half the trees because it was trained on North American forests, prompting manual annotation of 55,000 trees across 80 land parcels using drones and Meta's Detectron2
- Digital Green's FarmerChat app has reached more than 1 million farmers across South Asia and Africa, answering over 8 million questions in 16 local languages using small language models trained on 120,000 farmer queries built with agronomists
- Microsoft uses bioacoustics AI to monitor Amazon deforestation, while Brazil's Rare nonprofit converts real-time coastal data into WhatsApp voice alerts for fishers and oyster farmers in Pará state
- The ag-tech market was worth about $30 billion last year and is forecast to nearly triple to $84 billion by 2034, with Google, Microsoft, Amazon, IBM, and Alibaba all running AI agriculture programs
- The International Panel of Experts on Sustainable Food Systems warned that big tech partnerships with large agriculture firms risk focusing AI on corn, rice, wheat, soybeans, and potatoes — wrecking local food systems and deepening inequality
Why it matters: With about 2.3 billion people moderately or severely food insecure and the UN's zero-hunger goal likely to be missed, the $84 billion projected ag-tech market by 2034 is being shaped by tools that systematically fail smallholder farmers in the Global South — meaning the same communities AI is meant to help are instead being pushed into data-extraction relationships that mirror colonial extraction patterns, unless locally adapted alternatives like FarmerChat and Nakalembe's crop-mapping models scale first.
