Agricultural AI safety research / Research initiative

Agricultural AI should know when not to answer.

A fluent answer is not necessarily a safe or useful one. AgroGuardAI investigates how language models can provide more reliable agricultural guidance across different crops, languages and farming contexts.

Currently in research and experimentation
A farmer photographing crop leaves with a smartphone in a green field.
FIELD OBSERVATION Local knowledge. Real-world decisions.

The question behind the work

How do we make agricultural AI that is not only useful, but grounded, cautious and accountable?

We study model behaviour, the data it learns from and the decisions people may make based on its answers.

01 / Research focus

Fluency is not the same as reliability.

A model can sound certain while misunderstanding a crop, inventing a treatment or missing a critical safety precaution. Our research starts with these failure modes.

01

Factual grounding

Can an answer be traced to relevant, locally applicable agricultural evidence?

We examine unsupported claims, misleading diagnoses and the gap between fluent language and factual accuracy.

02

Language and context

Does the model understand the words farmers actually use?

We study local terminology, dialect variation and the farming context behind a question.

03

Safety and uncertainty

Can a model recognise when it should not give an instruction?

We investigate unsafe chemical advice, appropriate refusal and how systems communicate uncertainty when evidence is incomplete.

02 / The wider challenge

A global problem. No single-context answer.

Agriculture varies across regions, languages, crops and growing conditions. AI systems must account for that variation instead of treating farming knowledge as universal or giving confident answers without enough context.

01

Crop and field conditions — different crops, pests, diseases and growing environments.

02

Language and local knowledge — regional terminology, dialects and established farming practices.

03

Evidence and uncertainty — the quality of available information and what remains unknown.

04

Consequences of advice — risks to farmers, consumers and the surrounding environment.

03 / Research approach

Evidence before claims.

The work is organised around reproducible experiments and explicit limitations. Dataset size alone is not a result; performance must be evaluated against held-out questions and reviewed evidence.

I

Curate and document data

Record source provenance, language, crop context and risk category. Keep training examples separate from held-out evaluation data.

DATA
II

Run controlled model experiments

Study open-weight language models and fine-tuning approaches, recording model versions, configurations and dataset revisions.

EXPERIMENTS
III

Evaluate difficult and risky cases

Test grounding, language quality, unsafe advice, uncertainty and appropriate refusal across tasks and risk categories.

EVALUATION
IV

Seek independent review

Work towards review by qualified agronomists and native-language speakers before making claims about agricultural correctness or safety.

REVIEW

04 / Field notes

The research grew out of a practical failure.

Early farmer-facing prototype testing helped clarify the difference between making an AI assistant accessible and making its answers dependable.

Putting a general-purpose model inside an agricultural chatbot does not, by itself, make its advice reliable or safe.

Early versions of AgroMind were tested with more than 100 farmers. The experience exposed problems that a better interface alone could not solve: uncertain diagnoses, unsupported recommendations and the need for stronger local-language understanding.

That experience shifted the central question from how to make an agricultural chatbot answer more questions to how to evaluate whether its answers should be trusted.

Visit the AgroMind prototype

Research status

What this work has—and has not—established.

AgroGuardAI remains in the research and experimentation stage. The current dataset, training pipeline and screening diagnostics do not constitute a validated agricultural diagnostic or pesticide-prescription system. Automated checks cannot replace agronomic review, and the current evaluation tooling is not yet a publication-grade safety benchmark.

Experimental model outputs should not be used as the sole basis for chemical application or other safety-critical farming decisions.

Research collaboration

Help strengthen the evidence.

We welcome conversations with agronomists, native-language reviewers, AI safety researchers and research institutions interested in evaluating agricultural AI responsibly.

Contact the research team