I would not trust it. Considering the unfortunate critical mass of "bro-science", incorrect dispositions, and general statistical noise in the training data...you'd be better off just familiarizing yourself with an appropriate level tool, such as KIS Organic's 'A Dichotomous Key for Understanding Nutrient Deficiencies'.
If you insist on a tech-enabled solution you'd be better off using Google Lens and dynamically altering the scope of a high resolution image to investigate something you're on the fence for or lost about.
What is ChatGPT going to say when your leaves are showing an X deficiency but the real issue is actually over/underwatering, not the available nutrients? These are the boundary value cases where human intuition and experience comes in.
Now, can this be done? This absolutely could be done. We could do it together. We'd need a massive and CORRECTLY labelled dataset, we'd use an open-source imaging machine learning algorithm, and we'd iterated until it is the specified accuracy. There are similar/analagous applications of such models in industry right now, we can message more if interested.
On a personal note, being knee-deep in many technical topics professionally has led me to witness the sheer incompetence of nearly every LLM available to the market. Are you doing something where the answer is simply statistical bastardization of the finite nature of linguistics where data has already converged appropriately in the training? Then sure, ask that LLM do fix your grammar. Don't ask it about (say for example) legal advice or nuance IMO.