The growers are right about the chatbots

There is a lot of criticism that AI gives bad advice when it comes to growing cannabis — the internet cannabis forums are replete with comments about growers asking a question or uploading a picture of a plant, and AI giving back wrong, sometimes harmful advice. The message behind nearly all of the comments is that ChatGPT or Claude or Gemini just won’t work for growing. As a software developer who’s built an AI cannabis grow intelligence system, I just want to say: they are 100% correct that off-the-shelf generative AI can absolutely be wrong, sometimes spectacularly so.

So if I really am a developer of an AI cannabis grow system, aren’t I wasting my time? No.

What growers are saying

The complaints sort into four.

It is confident and wrong. It names a disease on a healthy plant. It recommends a light, at a distance, that would cook a seedling. It says both in the same assured voice it uses when it is right.

You cannot tell where it got the answer. Ask it for a source and you may land on a page selling the product it just recommended. Manufacturer claims, breeder claims, plant science and forum lore come out blended into one answer, with nothing to tell them apart.

It only answers what you ask. A new grower does not know which question to ask. The usual example is the grower who is sure he has a potassium deficiency. He does, in the narrow sense. The cause is nutrient lockout, and potassium is one of several things the plant cannot take up. A chatbot asked about potassium answers about potassium.

So they do not trust it. A grower who has been burned once discounts all of it. A grower who has never tried it says he does not know enough about it to trust it. Both are reasonable.

I agree with all four.

It is not wrong every time, and that makes it worse. Sometimes the advice is sound, and other growers will say so. A tool that is right one day and wrong the next, in the same voice, leaves you checking everything it says.

What an experienced grower does that a chatbot does not

Take the same question to an experienced grower and two things happen.

They ask questions back before they answer. Which leaves? What medium? What is your runoff pH? And they answer from what they know to be true about cannabis, not from everything anyone has ever typed about it.

Those are the two things the chat apps are missing.

1. A source. A fixed body of things known to be true about cannabis, so the answer is looked up from it and can say where it came from.
2. A procedure. The right questions, asked back, in an order, with the most likely hidden cause first.

Neither one is intelligence. A general chatbot has neither, so it composes something fluent. That is what it was built to do.

What it looks like with both

The box we build carries a knowledge pack for cannabis. One piece of it is the potassium case, written down.

When a photo reads as a probable deficiency, the dashboard does not name a nutrient. It offers a guided diagnosis, and the grower starts it. It opens with this, word for word:

The photo flags a probable nutrient deficiency, but several causes look nearly identical in a single image. The camera can’t see what tells them apart — your pH, runoff/EC, growing medium, and which leaves are affected.

Then it asks four questions:

1. Do you actually see discoloration or damage on the plant?
2. Which leaves are affected?
3. Besides the colour change, what are the leaves doing?
4. What is your last runoff pH, and what is your growing medium?

The rules run in a fixed order, and lockout is first. If the pH is outside the range for the medium (6.0 to 7.0 in soil, 5.5 to 6.5 in coco or hydro), the answer is lockout, before any nutrient is considered:

Your runoff pH is outside the uptake range for your medium, so the plant cannot absorb nutrients even if they are there — correct pH first before adding anything.

So a grower who arrives sure it is potassium, with a runoff pH out of range, is told to fix the pH first.

If the grower has no pH pen, it does not guess around the gap. It says: “Without a pH reading I cannot rule out lockout.”

The second rule sends the grower to numbers the box already has. Curling, clawing or crispy edges point to heat or humidity, not feed, and the hub measures leaf temperature and VPD.

Every one of those answers carries its source: chapter 5 of the Handbook of Cannabis Production in Controlled Environments (CRC Press, 2022). It was not composed. It was looked up.

IG-035A procedure and a source

What the guided diagnosis on the box does when a photo reads as a probable deficiency.

01 Four questions back Do you actually see discoloration or damage? Which leaves? Besides the colour, what are the leaves doing? Last runoff pH, and what medium?

02 Lockout first pH outside the range for the medium (6.0 – 7.0 soil; 5.5 – 6.5 coco or hydro): fix the pH before any nutrient is considered. No pen: 'Without a pH reading I cannot rule out lockout.'

03 Then the numbers the box has Curling, clawing or crispy edges point to heat or humidity, not feed — and the hub measures leaf temperature and VPD.

04 A family, with the page attached 'This is a likely cause, not a lab result.' Every answer cites chapter 5 of the Handbook of Cannabis Production in Controlled Environments. Looked up, not composed.

Offered only on a photo the classifier has read as a probable deficiency — about seven times in ten on real canopy pictures, and it once looked at a real problem 174 times and reported nothing.

Where a grower is still ahead

This is a short interview. It is not an hour with someone who has made the mistakes, and it is not a forum full of them.

• It names a family of causes, not an element. Its own closing line is “This is a likely cause, not a lab result.”
• It is only offered on a photo the classifier has read as a probable deficiency. On real canopy pictures our classifier is right about seven times in ten, and earlier this year it looked at a real problem 174 times and reported nothing. A grower can wait a long time for an offer that never comes.
• It knows what the hub measures and what the grower tells it. An experienced grower can look at the whole grow and ask about anything. Four questions cannot do that.
• A wrong answer in gives a wrong answer out, with a real citation on it.

A good forum will often settle a question like this in an afternoon, for free. Nothing here replaces that.

The narrow claim

The growers are right that a chatbot off the shelf is a bad grow advisor. What they are describing is a tool with no source and no procedure.

Both can be written down. Once they are, the machine’s job is small: ask the questions in order, measure what can be measured, and answer from the book with the page attached. That will not make it right every time. It does mean it cannot make up where the answer came from.

Sources

• Golem, S., Rootzone Management, ch. 5 §5.3, in Handbook of Cannabis Production in Controlled Environments, CRC Press, 2022. The source the interview’s answers cite.
• The interview text and pH ranges: our own cannabis pack, version 3.0.0. First-party.
• Seven in ten, and 174 looks with no finding: our own measurements, published in The AI said foxtailing.

Also read: The AI said foxtailing. Three growers said nothing was wrong. · Not a chatbot: what the assistant is allowed to say · The chatbot was never in the tent