Implications.
Specialized AI platforms trained on region-specific extension publications can offer more accurate and context-relevant advice to producers compared to generalized AI tools like ChatGPT.
By automating responses to common questions, generative AI may alleviate some of the workload for extension personnel, potentially reducing feelings of burnout and disconnect.
It is crucial for extension personnel and producers to validate AI-provided information with local experts to ensure its accuracy and applicability to specific contexts.
Increasing the effectiveness of AI tools in extension services requires active collaboration among state cooperative extension agencies to create a comprehensive, open-access repository of agricultural knowledge.
Introduction
The nexus of artificial intelligence (AI) and digital agriculture is expected to revolutionize agriculture as part of the “Fourth Industrial Revolution”. A vitally important subset of machine learning that has risen to prominence in the past several years is generative AI, which can generate new content in many formats including text, computer code, images, audio, and video (Khurana et al., 2023). Large language models (LLM) are a specific type of generative AI that deal exclusively with understanding and generating human language and are the key applications behind platforms such as ChatGPT (OpenAI, San Francisco, CA) and Gemini (Alphabet, Mountain View CA; Stokel-Walker and Van Noorden, 2023).
While useful for general questions, troubleshooting, and brainstorming, widely known LLMs such as OpenAI’s ChatGPT are inconsistent for domain-specific knowledge such as agriculture. The body of data used to train OpenAI models includes billions of pages of content, much of which doesn’t reach a high standard for reliability or accuracy. An easy-to-use, reliable tool that can effectively deliver digital information to extension personnel and agricultural producers is a critical need. This interface must offer the ability to iterate through a problem in a conversational format and do so in a simple narrative manner that is easily usable by the visually impaired. Recent developments in AI have made it possible to solve this need for producers and extension professionals. Here, we aim to explore the current uses and limitations of generative AI tools available for extension professionals and livestock producers.
Extension as a Resource
County Extension Agents (CEA) serve as educators and resources for the regional needs of both rural and urban populations for a wide range of topics. An extension office exists in nearly every county in the US. An agricultural CEA may work with crop producers, livestock producers, new landowners, minority producers, fruit growers, 4H members, and more. As a result, agricultural CEA often serve as generalists in terms of problem-solving. Although the average CEA has at least a Bachelor’s Degree, they are not equipped to be subject matter experts (SME) across multiple industries. Extension specialists, who are considered SME, are a primary source of technical support for CEA. However, specialists are spread thin; for example, most land grant universities only have a handful of livestock specialists on faculty. Additionally, both agents and specialists alike are expected to meet increasing reporting requirements, and petition for public and private grants as government funding generally decreases for extension services across the country. For a CEA, the lack of access to resources can lead to feelings of disconnect and burnout (Benge et al., 2015).
Aware of these shortcomings, the Extension Foundation with support from Oklahoma State University developed ExtensionBot, an agriculture and extension-specific generative AI platform that is designed with the express purpose of providing unbiased, science-backed knowledge from respected institutions. It taps into a collective body of literature (over 360,000 documents) authored primarily by research and extension professionals. The full list of contributing institutions can be found at https://extension.org/tools/extbot/. Many extension publications are written in response to commonly asked questions of livestock producers.
Case Study Comparison Between Tools
As the rate of land fragmentation increases in the United States, many new landowners are interested in owning livestock. A common question that agricultural CEAs receive from these new landowners is, “How many cows can I have on my land?” The true answer requires an assessment of how much forage the land can sustainably provide to cattle. Realistic thresholds vary based on ecoregion.
OpenAI currently offers 3 LLM: GPT-3.5 (training data cutoff September 2021), GPT-4 (training data cutoff April 2023), and GPT-4o (training data cutoff December 2023). To explore livestock-specific technical knowledge of those models, we built a function in RStudio (v 2023.06.2; R Core Group; Vienna, Austria) to ask each OpenAI LLM the same question 1000 times each. Interestingly, we received a different response each time. Table 1 shows the performance of the 3 LLM when asked if 5 acres is a sufficient stocking rate per cow/calf pair in either Potter County, Texas, or Jackson County, Florida. In Potter County, this would almost never be sufficient due to the semi-arid nature of the region; in Jackson County, greater rainfall and soil quality means 4 acres per cow/calf pair is usually sufficient (Mayo, 2022).
Table 1.
Comparison of responses of three large language models to a simple stocking rate question1
| Potter County, TX | Jackson County, FL | |||
|---|---|---|---|---|
| Model2 | Correct, % | Incorrect, % | Correct, % | Incorrect, % |
| GPT-3.5 Turbo | 85.9 | 14.1 | 15.9 | 84.1 |
| GPT-4 | 0.60 | 99.4 | 99.4 | 0.60 |
| GPT-4o | 99.8 | 0.20 | 0.50 | 99.5 |
1“Is 5 acres per cow/calf pair a sufficient stocking rate on unirrigated, continuously grazed land in (Potter County, TX/Jackson County, FL)?” For Potter County, TX, “No” was considered correct; for Jackson County, FL, “Yes” was considered correct.
2OpenAI models; iterated 1,000 times each.
After each response, the OpenAI models nearly always acknowledge the importance of context and implore the user to seek further guidance from a local county resource professional. Still, the inconsistency of OpenAI model responses is reflective of the many unknown sources the models utilize. Although OpenAI’s newer models will give links when prompted, those links are not necessarily what were used to generate the response. ChatGPT’s enormous training set of over 300 billion words combines with the lack of reliability factor to create frequent “hallucinations”—cases in which the model invents its own answers or URLs.
ExtensionBot will reply correctly (“yes”) to the Jackson County question because it is trained on University of Florida Extension publications. However, it does not correctly answer the Potter County question due to a lack of access to extension resources from that region. Its training set currently may not be encompassing enough to apply across all states and agricultural industries. As a result, the current ExtensionBot will be most useful to people in a state contributing to its training data.
The challenge lies in the need for active, open-access extension resources to develop platforms that can answer questions across regions. Greater collaboration and educational campaigns of generative AI tools amongst state cooperative extension agencies are needed to increase this repository. Tailored extension chatbots can be maintained and controlled by the extension services themselves, allowing incorrect answers to be flagged and changed as they arise. This is not a possibility with any other widely used chatbot platform. ExtensionBot is continuously fine-tuned to provide the most accurate answers to hundreds of questions commonly posed to CEA by comparing responses from alternative LLM. Currently, ExtensionBot uses Mistral 7b (Mistral AI; Paris, France) as its base LLM, but can implement different base models if others are found to provide more accurate answers.
Acceptance of Generative AI in Academia and Best Practices for Use
Academic professionals generally recognize the advantage of adapting to generative AI. The increased ubiquity of AI means it will be of utmost importance for extension professionals to understand the capabilities and limitations of chatbots. OpenAI’s LLM are free and fast. They are useful for generating outlines, note condensing, generalizing, and summarizing scientific topics in a mostly unbiased manner. They can answer specific scientific questions with relative accuracy, albeit without the nuance, real-world experience, and cautiousness required of a human SME (though, even human SMEs have a non-zero error rate). Extension chatbots are envisioned to serve as virtual consultants, reducing pressure on CEA and specialists by answering straightforward and commonly asked questions. Generative AI can also help CEA formulate
precise, probing questions for producers, enabling them to investigate vague or incomplete stakeholder inquiries more effectively. This could be leveraged into increased quality time between producers and extension agents to tackle more complex scenarios in an operation.
There are several limitations of generative AI as a virtual technical advisor for CEA. One is the static nature of the training data, which can quickly become outdated. Human connection is another limitation. Although AI can greatly enhance producers’ experience with the delivery of web-based extension content, one-on-one communication with a CEA remains important. Further, rural Americans often have limited access to broadband internet, leaving some producers skeptical of the applicability of digital solutions for their operations. Efforts to validate the accuracy of apps such as ExtensionBot while facilitating producer adaptation into the digital age are necessary.
Conclusion
Agriculture will need to adjust to the AI-enabled world like all facets of society. Stakeholders are advised to invest resources in tailored, agricultural-focused language applications to provide extension personnel and producers with more context-accurate answers than LLM like ChatGPT. Cooperation among land grants is encouraged to pool extension documents, though as pointed out above, site-specific factors can make it difficult to generalize answers across regions.
Acknowledgment
The views expressed in this publication are those of the author(s) and do not necessarily reflect the views or policies of the American Society of Animal Science, the journal, or the publisher.
Contributor Information
Jacquelyn Prestegaard-Wilson, Department of Animal Science; College Station, Texas A&M AgriLife Extension, TX, USA.
Jeffrey Vitale, Department of Agricultural Economics; Stillwater, Oklahoma State University, OK, USA.
About the Authors
Jacquelyn Prestegaard-Wilson, PhD, is an Assistant Professor and Extension Livestock Sustainability Specialist for Texas A&M AgriLife Extension. She develops Extension educational programs to increase the scientific literacy of livestock and environmental stewardship in Texas and across the United States. Her current programs focus on identifying barriers to sustainable grazing management strategies for cow/calf and stocker producers, the impact of nutrition on profitability, ration cost, and environmental footprint of beef and dairy cattle, and integration of time and cost-saving technologies for livestock producers. She is also the Texas A&M Department of Animal Science’s representative for the U.S. Roundtable for Sustainable Beef. Corresponding author:j.prestegaard@tamu.edu
Jeffrey Vitale, PhD, is an Associate Professor in the Department of Agricultural Economics at Oklahoma State University. He has been teaching an undergraduate farm and ranch management course for nearly 20 yr. His research interests include the adoption of new agricultural technology including GM crops, virtual fencing, and conservation tillage. He is currently part of a research team incorporating AI natural language applications to assist producers in accessing information in the digital age.
Conflict of interest statement
The authors declare no conflicts of interest.
Author Notes
These authors contributed equally to this work.
Literature Cited
- Benge, M., Harder A., and Goodwin J... 2015. Solutions to burnout and retention as perceived by county extension agents of the Colorado state university extension system. JHSE. 3(1):1. doi: https://doi.org/ 10.54718/NSXN7559 [DOI] [Google Scholar]
- Khurana, D., Koli A., Khatter K., and Singh S... 2023. Natural language processing: state of the art, current trends and challenges. Multimed Tools Appl. 82(3):3713–3744. doi: https://doi.org/ 10.1007/s11042-022-13428-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Mayo, D. 2022. Grazing Management – Making the most of the available forage. Gainesville (FL): UFIFAS Extension. [Google Scholar]
- Stokel-Walker, C., and Van Noorden R... 2023. What ChatGPT and generative AI mean for science. Nature 614(7947):214–216. doi: https://doi.org/ 10.1038/d41586-023-00340-6 [DOI] [PubMed] [Google Scholar]