We Are Not Robots
Everyone is talking about and using artificial intelligence or AI in different ways, including to help with plant problem diagnoses. Although AI is increasingly popular and can be a powerful tool, caution is warranted.
I decided to put this to the test. I took 10 photos from a random selection of past e-Gro Alerts, including pictures of disease symptoms, insects, symptoms of nutrient deficiencies, and symptoms of insect infestation. I uploaded each to ChatGPT and kept my inquiries very simple, asking "What is causing this?” or "What is this?”. Out of the 10 photos I challenged ChatGPT with, 3 were correct – it correctly identified the pictures of iron chlorosis on petunia, and spider mite damage on mini rose, and downy mildew on coleus. One of the 10 was somewhat correct, but mostly incorrect: for low substrate pH-induced iron toxicity on Christmas cactus, it incorrectly identified the problem as too much direct sunlight but correctly suggested a nutrient deficiency as a another possible cause. However, it mentioned incorrect nutrient deficiencies.
The majority were wrongly identified: tarnished plant bug was incorrectly identified as green aphid; downy mildew on creeping phlox was incorrectly identified as winter burn on rosemary; INSV on non-stop begonia was incorrectly identified as variegation; powdery mildew on sedum was incorrectly identified as fungal leaf spot, suggesting Cercospora or Alternaria; Chrysanthemum lace bug adults were incorrectly identified as immature stage whiteflies; phosphorous deficiency on osteospermum was incorrectly identified as spider mite damage.
Aside from the possibility of getting a diagnosis wrong, there are other important concerns: recommendations for management can be wrong or incomplete; products suggested may not be legally labeled or safe to use for the specific use, crop, or region; and recommendations will not be tailored for your situation. Sure, you might be able to get better answers with better prompts, but, in my opinion, the risk is too high.
I wondered what ChatGPT ‘itself’ would think, so I asked it: "Should I use chat GPT to diagnose greenhouse crop problems?” Its answer: "Short answer: You can use ChatGPT as a support tool — but not as your only diagnostic method. For greenhouse crop production, accuracy matters because misdiagnosis costs money fast.” …I think it got that one right.
Nora Catlin
Floriculture Specialist, Cornell Cooperative ExtensionNora Catlin is the Floriculture Specialist for Cornell Cooperative Extension of Suffolk County. She creates educational programs and materials, conducts applied research and demonstrations, and consults and advises growers on production, regulatory, or business issues.
