Why AI Is Ruining Your Grow

ChatGPT, Claude & Co. are not Plant Doctors: Why AI plant analyses without complete context are dangerously nonsensical and why growers need to learn to understand their plants themselves

A yellow leaf is photographed. Two images are uploaded to ChatGPT, Claude, Gemini, or another AI system. A sentence such as this is added:

“What is my plant lacking?”

A detailed answer appears just a few seconds later.

Calcium deficiency. Magnesium deficiency. Overwatering. Light stress. Incorrect pH. Nutrient lockout. Too little fertilizer. Too much fertilizer.

The AI explains the symptoms, names a presumed cause, and immediately provides an action plan. It recommends CalMag, flushing, a pH correction, more fertilizer, less water, or replacing the substrate.

The answer is clearly structured. It uses technical terms. It sounds logical and confident.

And it can still be complete bullshit.

That is why we deliberately state this clearly at CannaSelection:

General-purpose AI systems are currently not reliable tools for diagnosing cannabis plants. Anyone who carries out specific treatments based on two images is not performing a plant analysis. They are letting a statistical model guess based on incomplete information.

The danger is not only the wrong answer. The danger is that it sounds like expertise.

First, you need to understand what AI actually does

ChatGPT, Claude, and comparable systems are so-called generative foundation models. They are often simplistically referred to as large language models, or LLMs for short.

These models do not work like an encyclopedia in which a question is looked up and a verified expert answer is then issued.

During training, very large quantities of text, images, and other content are processed. The model learns statistical relationships, structures, and patterns from them. For example, it learns which terms frequently occur together, which explanations appear linguistically plausible, and how a particular type of question is usually answered.

OpenAI explains that its models are developed using, among other things, publicly available information from the internet, licensed content, and data provided or generated by people and users. Anthropic likewise describes a mixture of publicly accessible internet content, third-party data, internally generated data, and other datasets for Claude.

After training, however, the model does not store this content like a searchable database. It learns mathematical relationships within the training information and generates an answer that is likely to fit. OpenAI describes this process, in essence, as learning relationships within very large quantities of information.

Put simply:

AI does not know what your plant is lacking. It calculates which answer is likely to sound appropriate for your input.

That is impressively powerful for many tasks.

But it is something entirely different from a verified biological diagnosis.

Why a plausible answer is by no means necessarily a correct answer

A language model is trained to generate a meaningful continuation.

If someone asks:

“My cannabis plant has brown spots. What is it lacking?”

the model recognizes typical linguistic associations:

• brown spots,

• calcium,

• magnesium,

• pH,

• nutrient uptake,

• CalMag,

• overfertilization,

• light stress.

An answer emerges from these patterns.

However, the model does not automatically possess the information needed to distinguish between these possible causes.

It knows neither the root ball nor the water values. It does not know how the plant was watered over the past two weeks. It does not know whether the plant was recently repotted, whether CalMag has already been applied, or whether the presumed spots are mechanical damage.

Nevertheless, a generative model does not necessarily have to remain silent. It can replace missing connections with a plausible narrative.

The National Institute of Standards and Technology refers to this behavior as “confabulation”: generative AI systems can produce false content and present it with great confidence.

That is precisely what makes AI plant analyses so problematic for beginners.

An experienced grower may recognize that a recommendation does not fit the substrate or the cultivation practices used so far.

A beginner often only recognizes this:

“The AI said it was calcium.”

And then calcium is applied.

Image-recognition AI does not understand a plant either

Modern systems such as ChatGPT or Claude can process images as well as text. This quickly creates the impression that they assess a plant photo similarly to an experienced plant pathologist.

That is wrong.

Multimodal models process visual features and connect them with linguistic patterns. They can recognize that a leaf is yellowish, has spots, is curling, or has a particular shape.

That does not mean, however, that they automatically understand which biological process caused this appearance.

A leaf can look similar in the case of:

• an actual nutrient deficiency,

• impaired nutrient uptake,

• an overloaded root zone,

• substrate that is continuously too wet,

• an excessively high salt concentration,

• unfavorable transpiration,

• light or heat stress,

• pest infestation,

• mechanical damage,

• natural aging,

• genetic expression.

The image shows the visible consequence.

It does not necessarily show the cause.

Even models specifically developed to detect plant diseases have considerable difficulty when real-world images differ from the conditions represented in their training datasets. Lighting, background, camera angle, cultivar, developmental stage, and environment can significantly affect reliability. In research, this problem is known as domain shift.

These are models developed specifically using labeled plant images.

ChatGPT, Claude, and other general-purpose AI systems, by contrast, are not validated cannabis diagnostic models.

The ability to describe an image is not the same as the ability to diagnose a plant.

Cannabis is a particularly poor use case for AI

Cannabis was prohibited or strictly regulated in many countries for decades.

This had significant effects on research, data quality, and the open documentation of professional cultivation.

Regulatory barriers have demonstrably made cannabis research more difficult. Scientific institutions cite limited access to suitable plant material, complex approval procedures, limited funding, and a lack of standardization as key problems, among others.

A large proportion of existing cannabis research concerns medicinal effects, active compounds, consumption, and health consequences.

AI-assisted diagnosis of indoor cannabis plants, however, would require entirely different datasets.

A robust diagnostic model would need to link images with extensive and verified supporting information:

• genetics,

• plant age,

• developmental stage,

• substrate composition,

• pot or bed volume,

• root condition,

• water analysis,

• watering history,

• fertilization strategy,

• temperature profile,

• humidity,

• leaf and room temperature,

• PPFD and DLI,

• pest status,

• laboratory analyses,

• confirmed cause,

• documented treatment,

• actual response to that treatment.

Thousands of comparable cases would have to be documented under controlled conditions, assessed by experts, and monitored over the long term.

Precisely this data basis does not exist for cannabis on the scale and with the quality required for reliable general image-based diagnosis.

Instead, a model will often find the following on the internet:

• forum posts,

• Reddit discussions,

• grow journals,

• manufacturer texts,

• product advertising,

• blog articles,

• social media posts,

• incorrectly labeled symptoms,

• personal individual experiences,

• deficiency charts copied from one another.

This is not a scientifically validated diagnostic database.

Yes, Reddit and forums are also part of the information environment

We need to be aware of what type of information is available on the public internet.

OpenAI confirms that publicly available internet content is among the data sources used. Since May 2024, there has also been an official partnership between OpenAI and Reddit through which OpenAI can access structured and current Reddit content.

This does not mean that every ChatGPT answer comes directly from a specific Reddit comment. A model does not usually simply copy an individual post and cite it as a source.

It does, however, mean that public discussions, community opinions, and unverified user content can be part of the information environment from which such systems learn patterns or obtain current information.

And especially with grow-related topics, we need to examine the quality of this starting point.

In many forums, plant problems are diagnosed without knowing the water values, watering history, or substrate composition. One user writes “CalMag,” the next “nitrogen,” and the third “overwatered.”

No one examines the roots.

No one performs a substrate analysis.

No one reliably documents which measure actually helped in the end.

Nevertheless, these posts remain publicly available and continue to be cited, copied, and summarized.

If the source information consists of speculation, half-knowledge, and incorrectly labeled images, it cannot produce reliable cannabis diagnostics.

A model can generate a linguistically excellent summary from a large quantity of poor information.

That does not automatically make the information true.

The person before the AI is often the second problem

Even a very powerful model can only work with the information it receives.

And this is precisely where the next problem begins.

Most growers do not give the AI the complete cultivation history. They pass on what they themselves perceive.

These are two completely different things.

For example, a beginner writes:

“The plant is turning yellow even though I water normally.”

What does “normally” mean?

• How many liters?

• At what intervals?

• In what pot size?

• What criterion is used to decide when to water?

• Is there runoff?

• How quickly does the substrate dry?

• How extensively is it rooted?

• Is the entire surface watered evenly?

• Is the substrate at the bottom permanently wet?

• How heavy is the pot before and after watering?

The user may believe they are watering correctly. That is why they do not mention their watering practices as a possible source of error.

Consequently, the AI does not receive the actual facts.

It receives the user’s subjective interpretation.

The same applies to statements such as:

• “The climate is fine.”

• “The pH is correct.”

• “The light is not too strong.”

• “The soil is high quality.”

• “I only fertilize lightly.”

• “The plant cannot be overwatered.”

• “It must be a deficiency.”

The AI does not see the grow.

It sees the grower’s account of their grow.

If important information is missing, the model cannot conjure it out of the image.

Instead, it may fill the gaps with typical assumptions.

More context can improve the answer, but cannot guarantee a diagnosis

Of course, AI can produce better results when it receives extensive and precise information.

Anyone who fully documents genetics, plant age, substrate, pot volume, water analysis, lighting, climate, watering quantities, fertilization, and the timeline enables a much better assessment than someone who uploads only two images.

AI can then, for example:

• structure possible causes,

• identify contradictions in the information provided,

• formulate useful follow-up questions,

• develop a differential diagnosis,

• compare measurements,

• summarize documentation.

However, this is not a justification for blind trust.

Even with more context, a general-purpose model remains an unvalidated cannabis diagnostician. It can set incorrect priorities, mix up different cultivation systems, or develop a logically plausible but incorrect conclusion from a false user statement.

The crucial distinction is:

With good context, AI can support the thinking process. But it should not take over the decision.

This is exactly where typical use fails.

Most people do not enter all relevant information. They upload a photo, expect a definitive answer, and then act as though a laboratory analysis had been performed.

A plant is a system, not a collection of individual spots

Cannabis plants respond to the interaction of numerous factors.

Water influences oxygen levels in the root zone.

Root condition influences nutrient uptake.

Temperature and humidity influence transpiration.

Transpiration influences the transport of certain nutrients.

Light intensity changes water and nutrient requirements.

Substrate structure determines how water and oxygen are distributed in the pot.

A visible symptom can therefore stand at the end of a long chain of causes.

For example, a plant can show signs of a deficiency even though the relevant nutrient is present in sufficient quantities in the substrate.

The cause may instead be:

1. The substrate remains too wet for too long.

2. The root zone receives too little oxygen.

3. Root function is impaired.

4. Nutrient uptake deteriorates.

5. Apparent deficiency symptoms develop on the leaves.

The AI recognizes the symptoms and recommends additional fertilizer.

This leads to further watering, increases the salt concentration, and exacerbates the actual problem.

The visible observation was correct.

The interpretation was wrong.

This is particularly dangerous in Living Soil

Living Soil is a biologically active system.

Nutrient availability, microbial activity, organic matter, moisture, oxygen supply, roots, and substrate structure are directly interconnected.

Nevertheless, we regularly encounter AI recommendations such as:

• flush Living Soil thoroughly,

• immediately increase the EC,

• apply CalMag as a blanket response to spots,

• add mineral fertilizer according to a schedule,

• aggressively correct the pH of every batch of irrigation water,

• combine multiple organic and mineral products,

• immediately treat presumed deficiencies with individual nutrients.

Some of these measures may be useful in specific situations.

Without understanding the entire system, however, they are not a diagnosis.

They are actionism.

It is particularly problematic that AI frequently combines information from mineral, organic, and hydroponic systems. The answer may contain individual correct statements from each area and still be completely wrong as an overall recommendation.

This creates a kind of technical hybrid product:

• a little hydroponics,

• a little organic cultivation,

• a little Living Soil,

• a little manufacturer description,

• a little Reddit.

Everything fits together linguistically.

Biologically, it makes no sense.

AI can quickly turn a minor problem into a fail grow

A cannabis plant needs time to respond to changes.

A leaf that has already been damaged will not necessarily turn green again after a correction. A symptom may originate from a problem that occurred several days earlier. Some irregularities may already have stabilized and only need to be monitored.

AI answers, however, almost always end with a recommended action.

This creates the feeling that something must be done immediately.

The typical sequence then looks like this:

1. The AI identifies a presumed calcium deficiency.

2. CalMag is applied.

3. After two days, the old leaf looks unchanged.

4. A second AI suspects overfertilization.

5. The substrate is flushed.

6. A pH problem is then suspected.

7. The pH is corrected.

8. Fertilizer is added again afterward.

Within a few days, every relevant variable has been changed.

Afterward, no one can determine:

• what originally happened,

• which measure was useful,

• which measure caused new damage,

• whether any intervention was necessary at all.

A significant proportion of failed grows are not caused by the initial problem, but by the frantic reaction to it.

AI accelerates this actionism.

If you only ask AI, you will never get to know your plant

The larger problem is not technical.

It is human.

Since legalization, we at CannaSelection have increasingly observed that growers no longer provide a cultivation history.

They send two images and expect a finished solution.

What is missing includes:

• genetics,

• age,

• pot volume,

• substrate,

• water values,

• light values,

• climate,

• watering quantity,

• fertilization,

• development over time,

• measures already carried out.

Follow-up questions are often met with confusion.

Not because this information is deliberately withheld, but because many growers have never engaged with it.

They do not know exactly how much water they provide.

They do not know their water values.

They cannot explain why they use a product.

They have not documented the progression.

They do not observe which leaves were affected first.

They do not want to understand a plant.

They want an answer.

But that is not how cultivation works.

Growing means understanding the why

Anyone cultivating a plant should be able to explain the following for every measure:

• Why am I using this substrate?

• How does it retain and distribute water?

• How does oxygen reach the roots?

• Where do the nutrients come from?

• How do they become available to the plant?

• Why am I watering today?

• Why am I providing this quantity of water?

• What is the objective of this input?

• How do I recognize whether the measure is working?

• What side effects can it have?

If someone cannot answer these questions, they should not implement the next AI recommendation.

They should first understand their own system.

AI can formulate a technically convincing justification for almost any idea.

But good wording does not make a bad measure sensible.

If you enter bullshit, provide incomplete information, or do not know the decisive factors yourself, you will often receive nothing more than better-worded bullshit in return.

Focus on plants instead of prompts

Anyone who wants to grow reliably needs to engage with the fundamentals:

• plant physiology,

• plant symptoms,

• root health,

• irrigation,

• substrate structure,

• soil biology,

• light,

• transpiration,

• temperature,

• humidity,

• water chemistry,

• nutrient cycles.

This sounds more time-consuming than uploading a photo.

But it is the only way to make good decisions consistently.

A grower does not need to memorize every biochemical reaction. But they must understand that a symptom can have several causes and that a plant must never be considered independently of its system.

They need to learn to read their plant:

• How is the leaf positioned?

• How does it feel?

• How is the new growth developing?

• How quickly does the plant drink?

• How does it respond after watering?

• Is the problem static, or is it spreading?

• Does it affect old or young leaves?

• Is only one plant affected, or the entire crop?

This information does not come from a prompt.

It comes from daily observation.

Our clear recommendation

We expressly advise against using ChatGPT, Claude, Gemini, or other general-purpose AI systems as a basis for diagnosing and treating cannabis plants.

In particular, no one should, based on an AI response:

• immediately add fertilizer,

• flush Living Soil,

• apply several inputs simultaneously,

• mix organic and mineral strategies,

• change the pH hastily,

• completely change functioning processes,

• prioritize individual measurements over the plant.

AI cannot lift your pot.

It cannot feel whether the substrate is compacted.

It cannot smell an anaerobic root zone.

It does not know how the plant developed over the past few days.

It does not see how the plant behaves during the dark period.

It only knows what you tell it and what can be seen in a limited image section.

It does not know your grow. Therefore, it should not control it either.

Why the Grow Doctor works differently

The Grow Doctor by CannaSelection was deliberately not designed as a supposed AI Plant Doctor.

It is not intended to invent a spectacular diagnosis from two images.

Instead, it guides growers systematically through symptoms and relevant influencing factors.

Among other things, it narrows down:

• which part of the plant is affected,

• whether old or young leaves are showing symptoms,

• how the symptoms are structured,

• which causes are generally possible,

• which conditions need to be checked additionally,

• which causes can be systematically ruled out.

That is the crucial difference.

The Grow Doctor does not promise a magical instant answer.

It encourages the grower to look more closely and check the right variables.

Because plant diagnostics are not a picture quiz.

They are a process of elimination.

Back to the Roots

Good home cultivation does not begin with ChatGPT, Claude, and Co.

It begins with curiosity, observation, and a willingness to take responsibility for a living system.

Anyone who wants only a finished product but is unwilling to understand processes or invest time in their plants should honestly question whether home cultivation is the right project for them at all.

Provided the legal and medical requirements for treatment are met, going through a pharmacy may ultimately be more affordable, controllable, and reliable in quality than a poorly understood indoor grow.

And we say this even though, from our perspective, standardized pharmacy quality should not be the goal of home cultivation, but at most its lowest benchmark for comparison.

Home cultivation can achieve considerably more.

But only if the grower is willing to engage with the plant.

Our conclusion

ChatGPT, Claude, and other generative models are impressive tools.

They can structure information, prepare texts, compare measurements, and, when provided with complete context, organize possible hypotheses.

But they are not cannabis Plant Doctors!

They were not trained on a comprehensive, scientifically validated database of cannabis symptoms. They know neither your individual setup nor your cultivation history. Their answers are based on statistical patterns from very different sources—including potentially forums, community discussions, manufacturer content, and incorrectly classified symptoms.

The result can sound plausible and still be completely wrong.

That is why our advice is:

Stop having your plants diagnosed by general-purpose AI systems.

Get to know your plants.

Understand your soil.

Know your water.

Document your measures.

Observe changes.

And question why you are doing something every time.

A grow does not improve because an answer appears within ten seconds.

It improves when the grower understands what is actually happening in their system.