Hallucination Behavior in Multimodal LLMs Across Agricultural Image Interpretation and Generation Tasks

A comprehensive study reveals that multimodal large language models exhibit significant hallucination problems in agricultural imaging tasks, with image interpretation achieving only 63-75% zero-shot accuracy and text-to-image generation producing up to 91% biologically inconsistent scenes. These findings highlight critical reliability gaps that could undermine the trustworthiness of AI-driven agricultural platforms.

This research exposes a fundamental vulnerability in deploying multimodal LLMs for domain-critical agricultural applications. The study systematically evaluates hallucination patterns across two task types: interpreting crop disease and environmental stress from images, and generating synthetic agricultural scenes from text prompts. The modest baseline accuracy rates—ranging from 63-75% in zero-shot image interpretation—demonstrate that current models lack robust visual reasoning capabilities when domain expertise is required.

The findings reflect a broader challenge in AI development: models trained on general internet data struggle with specialized knowledge domains where accuracy directly impacts economic and food security outcomes. Few-shot prompting improved interpretation accuracy to 86.8%, but residual hallucinations persist, suggesting that prompt engineering alone cannot overcome architectural limitations. The text-to-image results are particularly concerning, with advanced models like GPT-4 and Gemini 2.5 Flash generating biologically implausible agricultural scenes in 91% of cases under relaxed constraints, revealing that generative models lack fundamental understanding of agricultural biology.

For agricultural stakeholders and agtech companies, these results indicate that deploying LLM-based imaging systems without rigorous validation creates significant risk. Misidentified crop diseases could lead to inappropriate pesticide use or missed interventions, with cascading economic and environmental consequences. The research underscores the necessity of human expert oversight and domain-specific fine-tuning before agricultural AI tools reach farmers. Moving forward, developers should prioritize domain-informed evaluation metrics and hybrid approaches combining LLM reasoning with specialized agricultural models rather than relying on general-purpose multimodal systems.

• →Multimodal LLMs achieve only 63-75% accuracy in zero-shot agricultural image interpretation, with significant hallucination rates affecting disease and stress detection.
• →Few-shot prompting improves accuracy to 86.8%, but hallucinations persist, indicating fundamental model limitations beyond prompt engineering solutions.
• →Text-to-image models generate biologically inconsistent agricultural scenes in up to 91% of cases, revealing deep gaps in biological understanding.
• →Deploying unvalidated LLM-based agricultural platforms poses real risks to farming decisions, crop health management, and food security.
• →Domain-specific fine-tuning and hybrid AI approaches are essential before agricultural LLMs can be reliably deployed in production environments.