Why general-purpose AI falls short in agriculture
Reported by TechRadar1 source ↓
TechRadar reports that general-purpose chatbots gave unsafe or mistimed crop-treatment advice in reported cases and research tests, arguing that useful farm guidance needs field-specific data and agronomist input.
Key points
- Researchers found ambiguous pesticide advice from GPT-3.5 and GPT-4 in tests involving fall armyworm
- ChatGPT gave Nigerian cassava farmers simulated advice with incorrect herbicide timing
- A McKinsey survey cited by TechRadar found about 17% of farmers worldwide use generative AI for daily decisions
Details
TechRadar describes a farmer in Chuzhou, China, whose sesame seedlings died after he sprayed a chatbot-recommended mix across 25 acres. The reported loss was about 150,000 yuan, or roughly 18,000 euros. The chatbot later acknowledged errors in its recommendations, including the choice of treatment and advising a whole-field application.
The article also describes research published in Nature Food that tested generative AI as a possible substitute for human agricultural advisers. GPT-3.5 and GPT-4 gave ambiguous pesticide advice for fall armyworm, while simulated questions from Nigerian cassava farmers elicited herbicide advice with incorrect application timing. The article says following that timing would have killed the crops.
TechRadar argues that general-purpose chatbots lack field conditions such as crop growth stage, local weather and the location of weeds, and may not ask for missing information before giving advice. It says agricultural models could use field data and guidance from agronomists to tailor recommendations and set limits on unsafe advice. The article also cites a McKinsey survey finding that about 17% of farmers worldwide already use generative AI in day-to-day decisions.
Written from coverage as of
Images
- TechRadar