Trust in Agricultural AI “Begins with Reliable Data”…Seeking Verification and Accountability Solutions at the Korean Society for Agricultural Machinery Fall Academic Conference

Among 25 greenhouse tomato farms in North Jeolla that have used integrated environmental control for more than 10 years, not a single farm answered that “environmental stability improved the most.” None of the farms considered maintenance and technical support when adopting the system. Yet among the six complaints that emerged during actual use, four concerned environmental stability. As for technologies needed for wider adoption (multiple responses), “an easy-to-use screen” and “remote monitoring and diagnosis” each accounted for 40% (10 farms), while “AI decision-making support” was only 24% (6 farms).

The survey was conducted on the 16th of this month by Rose Valley CEO Jung Byung-gyu among 25 graduates of the 2016 tomato course at Jeonbuk Agricultural Meister College, and was presented at the special session “Trustworthy Agricultural AI, From Principles to the Field” at the 2026 Fall Academic Conference held by the Korean Society for Agricultural Machinery on the 21st. At the session held that day at BEXCO in Busan, CEO Jung, Jeon Chang-bae, chairman of the International Association for Artificial Intelligence Ethics, Park Heun-dong, CEO of Jinong Co., Ltd., and Yang Seung-yeop, research fellow at the Korea Labor Institute, gave presentations. In the ensuing panel discussion, moderated by Kim Min-young, director of the Technology Convergence Strategy Division at the Rural Development Administration, the presenters were joined by Professor Moon Gwang-jin of Mokpo National University, Jeong Hwi-woong, director at 42MARU, and Jeong Hwan-seok, deputy director for computing at the Rural Development Administration. Participants discussed the risk factors of agricultural AI, verification methods, accident liability, and necessary norms.

If Sensors and Data Waver, Judgments Cannot Be Trusted…On-Site Risks Are Equipment and Data

CEO Jung, who presented the survey, operates a 1㏊ tomato greenhouse in Iksan, North Jeolla. After introducing a smart farm system, he reduced daily management time from 12 hours to 6 hours and raised yield per 3.3㎡ from 80㎏ to as high as 142㎏. Jung explained that this achievement was “not an effect created by AI alone, but the result of equipment, data, and the management of people who observe the field every day working together.”

He also introduced specific cases of malfunction. Incorrect sensor measurements disrupted ventilation window control, allowing mold to spread, and incorrect nutrient solution electrical conductivity (EC) settings damaged roots. At one farm, windows remained closed even after a shower had stopped, causing the temperature inside the greenhouse to exceed 60℃, and the grower ultimately abandoned that cropping season. Another farm could not receive parts for an imported controller in time and had to control the system manually all night. Jung said, “If sensors and data are unstable, no matter how smart the judgments built on top of them are, they cannot be trusted,” adding, “Trustworthy agricultural AI begins with trustworthy data.”

In the panel discussion, Director Jeong Hwi-woong said accuracy should be viewed by dividing it into “measurement” and “execution.” If execution goes wrong, he said, it is also necessary to determine whether the fault lies with the AI or the machine. Jeong added, “Data is not something you build once; it must be managed continuously.” Deputy Director Jeong, who is in charge of the AI farming assistant “Isak-i” at the Rural Development Administration, said, “There is not much data suitable for AI learning,” identifying data standardization and refinement as priority tasks.

Decision-Support AI…Final Judgment Should Remain With Humans

As the first direction for improvement, Jung proposed that “AI should move toward decision support that helps judgment rather than replacing humans.” The remaining directions are the standardization of sensor specifications, localization and rapid after-sales service, farmer education, and climate-adaptive control.

CEO Park emphasized that AI that explains the grounds for its judgments is a better service. Park said, “It would be good to move toward fully unmanned automation, but if it goes wrong, farms could become foolish,” adding, “The ultimate goal of AI is to keep producing farms at the level of CEO Jung.” Deputy Director Jeong also said, “For now, the safest approach is to use AI answers as a reference, checking and confirming them, rather than trusting them 100%.”

“Who Will Believe These Numbers?”…Calls for Independent Verification and Certification Systems

CEO Park trained a speech recognition model using recordings of farm consultations. In its own evaluation, the accuracy was higher than that of Naver Clova. However, Park said, “The problem is who will believe these numbers,” adding, “Right now, we are making the questions and doing all the evaluation ourselves.” As an alternative, he proposed independent benchmarks. The idea is for academic societies and the Rural Development Administration to set tests by area—speech recognition, evidence-based answers, pest and disease diagnosis, and prediction and recommendation—and link the results to a grading and certification system.

Director Jeong offered a different view. Since speech recognition will catch up with the highest level within two to three years, he suggested narrowing the scope of evaluation and focusing on standards for determining damage caused by prediction errors. Research Fellow Yang cited the aviation industry as an example and said, “Providing a sense of stability through regulation actually helps innovation and diffusion.” He also proposed giving public procurement incentives to companies that receive reliability certification. CEO Jung emphasized open field demonstrations in which farmers participate from the early stages of development.

If an Accident Happens, Who Is Responsible?…Decision Histories Should Be Kept and the Burden of Proof Reduced

In his presentation, Chairman Jeon introduced accidents that have already occurred in agricultural settings. These included an accident in November 2023 at a paprika sorting facility in Goseong, South Gyeongsang, in which a robot mistook an employee of a company who was inspecting it during a trial run for a box and killed him, and an accident in which a farmer’s finger was severed during drone-based pest control. Jeon said, “AI safety is a matter of prevention, not resolution.” Deputy Director Jeong said the sorting facility accident appeared closer to an issue of exception handling in the program, and said safety verification should be incorporated from the research and development stage.

CEO Park proposed a history-tracking system that records AI recommendations, farmers’ approvals, and execution results by time point. The idea is to make it possible to trace back where an error occurred if an accident happens. Research Fellow Yang said procedures should be established to divide responsibility like a ratio of fault, rather than concentrating it on one actor. Yang said, “It will become an important issue to shift or ease the burden of proof so that it is not farmers who prove problems with AI, but the companies that created the AI who prove otherwise.”

After the AI Framework Act, an Agricultural AI Special Act?…“In the Short to Medium Term, Supplement the Smart Agriculture Act”

Chairman Jeon said the AI Framework Act, which took effect in January this year, is a declarative and comprehensive law, and proposed that an “Agricultural AI Special Act” is needed. He also pointed out that when autonomous agricultural machinery causes an accident on the road, the applicable law is “currently in a vacuum.” Professor Moon said, “There are not many things in agriculture that can be regarded as high-impact AI, so a special act is a long-term task,” adding, “In the short to medium term, we should start by supplementing the Smart Agriculture Act.”

Research Fellow Yang unveiled a draft of the “Agricultural AI Ethics Guidelines” being prepared through research commissioned by the Rural Development Administration. In a preliminary survey, 16 out of 19 experts said guidelines were needed. The draft presented six core principles: ▲respect for humans ▲fairness and inclusion ▲safety ▲transparency ▲privacy ▲accountability. Whether to broaden the scope of protection to the overall ecology and environment, including soil and water quality, and how to address compensation for farmer data are still under discussion.

Education and Public Forums Needed to Reduce the Digital Divide for Elderly and Small-Scale Farmers

Research Fellow Yang emphasized the need to support AI education and access for digitally vulnerable groups, including elderly and small-scale farmers. Professor Moon said an agricultural AI public forum is needed, citing the absence of agriculture-related institutions or experts on the National AI Strategy Committee. Chairman Jeon said education, technology development, and the improvement of laws and systems must proceed together. Deputy Director Jeong said he would review the formation of a consultative body to expand the use of agricultural AI.

Director Kim, who chaired the discussion, concluded the debate by saying, “Securing the reliability and safety of agricultural AI does not restrict technological innovation; it is the foundation for diffusion in the field.”