A Handheld Extension Agent? GROW Researchers Put ExtensionBot to the Test

AI is sowing new possibilities in agriculture, from precision herbicide applications to mapping weeds and cover crops in fields. Now meet ExtensionBot, a large language model chatbot created by The Extension Foundation that answers agricultural questions by pulling information from a pool of universities and public institutions

“From the responses I saw, farmers will still need to make their own decisions, but AI may be able to help synthesize information faster and more efficiently.”

Dr. Michael Flessner, Virginia Tech

But how sharp is this virtual Extension agent? GROW researchers put ExtensionBot to the test, discovering its limits. While ExtensionBot tackled general weed management questions easily, Virginia Tech’s Dr. Michael Flessner and Penn State University Emeritus Dr. Bill Curran found it struggled with questions that required nuanced or complex answers, beyond the basics of a topic. 

“When I didn’t like an answer, it was usually because of a nuance, or the answer wasn’t fully complete,” explains Flessner. “For example, ExtensionBot hit most every aspect of maximizing the efficacy of Liberty herbicide, but only indicated sunny conditions were needed and didn’t delve into time of day”, which can also be important to maximize Liberty performance.  

The researchers also found that ExtensionBot might struggle staying current with product labeling. For example, ExtensionBot recommended postemergence dicamba for use in soybeans, despite postemergent dicamba products being now unavailable.

Overall, the bot stumbled most often when tackling detailed questions about herbicide selection and management, suggesting that even AI struggles to digest herbicide labels – a struggle that should surely endear the robotic Extension agent to many a crop consultant and farmer. 

On the plus side, ExtensionBot’s detailed list of sources for its answers help to clarify whether the bot correctly interprets its findings. Flessner used these sources to check the validity of the bot’s response himself. “Verification is still important,” he says. “In instances when I felt an answer was wrong, it was easy to diagnose by clicking on the source and figure that out using ExtensionBot.”

The table below lays out a sampling of the questions the GROW researchers sent to ExtensionBot. While these queries span a variety of subjects, they don’t encompass every topic that a farmer or Extension agent might have. 

“I didn’t push the bot for detailed, nuanced recommendations, that consider a farmer’s crop rotation, soil type, weeds, tillage practices, planting dates, or equipment limitations.” Flessner explains, “From the responses I saw, farmers will still need to make their own decisions, but AI may be able to help synthesize information faster and more efficiently.”

With ExtensionBot still being developed, there is potential for the chatbot to continue improving and eventually offer more nuanced and accurate information. Its current state already makes it a handy tool for those looking to get answers for general questions. And ExtensionBot will be but one of the possible AI sources out there, with Google’s AI overview, ChatGPT and Grok, X’s AI chat feature, providing very similar services, the researchers noted. 

“It could be very useful for county agents or others that need quick answers to more general questions.” Curran says.

So dig in to some of the questions the researchers posed to this handheld Extension agent, and see how it fared. Better yet, take your own turn chatting with ExtensionBot, and let us know what you think! 

ExtensionBot’s performance on sample questions, as ranked by some GROW Extension researchers. (Chart credit: Emily Unglesbee, GROW)

Article by Emily Unglesbee and Amy Sullivan, GROW; Michael Flessner, Virginia Tech; and Bill Curran, Penn State Emeritus.