Using AI in agriculture can enable quick access to information, but it proved costly for a Chinese farmer. A faulty recommendation from an AI application destroyed his sesame crop on about 25 hectares of land. The incident highlights the limits of relying on AI models when their responses move from the screen to real-world decisions and the consequences of errors are immediate and expensive.
Gradual trust in an error-prone tool
A 67-year-old farmer known only as "Wu" had been using an AI app for guidance on farming and dealing with pests and weeds. After several successful experiences, his trust in the tool grew and, ultimately, without consulting an expert or testing on a small scale, he implemented the AI's recommendation to combat weeds and pests across a large area of his sesame field.
A herbicide that also destroyed the crop
The app had suggested a mix of chemicals including a herbicide for broadleaf plants, unaware that sesame is also a broadleaf plant. As a result, instead of protecting the crop, the substance damaged it. One day after spraying, the sesame seedlings wilted and a large portion of the field was lost. The damage was estimated at about 25 hectares (150 mu) and its monetary value was reported at close to 150,000 yuan (about 20,700 dollars).
Why do language models make mistakes?
Large language models are essentially designed to produce coherent linguistic responses, not to make specialized decisions in real agricultural conditions; this can lead to confident but incorrect answers. Research published in July 2024 in Frontiers in Plant Science shows that general models are limited in adapting to local agricultural conditions and a single model cannot address all environments and tasks. Correct agricultural recommendations require data about crop type, pest, soil and weather conditions, and permitted pesticides; therefore, correctness of a response in one situation does not guarantee correctness in another.
Prior warnings about selecting pesticides with AI
This is not an isolated case; Cornell University previously warned against unverified reliance on AI recommendations about pesticides, because some systems may suggest substances that are not permitted for the intended crop. Off-label use can harm the crop, the environment, and people's health; therefore checking the label and consulting a specialist before implementation is essential.
Not a replacement for an agricultural expert
The incident does not mean AI is useless in agriculture; applications such as pest detection, image analysis, and irrigation need estimation are still expanding. Researchers suggest using specialized models tied to authoritative databases and information-retrieval–based decision support systems that consider local conditions, rather than general models.
A lesson beyond agriculture
The key point is the gap between a persuasive answer and a correct, actionable one. The riskier the recommended decision, the more important human review and verification become; because with pesticides, a small mistake can destroy a large crop within hours. AI is a useful tool, but it should not be the final link in the decision-making chain.