ChatGPT for Farmers: Can It Give Reliable Farming Advice?

Not reliably, and not for anything where a wrong number costs a crop, an animal or a licence. ChatGPT can give fluent advice on crops and livestock that is wrong on timing, rates or product choice, so doses, spray rates and withdrawal periods must come from the label and your adviser. It is dependable for admin, drafts and tidying records you supply.

The reason is how it works. It writes what a good answer would probably look like, based on patterns in text, and it does so with the same confident tone whether it is right or wrong. A farmer cannot tell the difference from the wording, and a mistake on spray timing shows up months later in a yield figure, not in a typo you can spot.

Follow me on Instagram@sagnikteaches

The split is easy to apply once you see it. "Draft a polite chaser to the seed merchant about a late order" is a safe request, because you will read every word before it goes. "What rate should I spray on my stubble?" is not, because the number belongs to the product label, and a fluent wrong answer costs a field.

Connect on LinkedInSagnik Bhattacharya

What has actually gone wrong, from published reports

I looked for specific, documented cases rather than general warnings. The evidence is thinner than the opinions, and mostly dates from the first versions of these tools, so read it as a pattern, not a current score.

Subscribe on YouTube@sagnikteaches

• Herbicide timing and planting advice. Ambrook's August 2024 piece on whether chatbots are worth the risk (the site now publishes as Offrange) reports a study in Nature Food of ChatGPT's advice to cassava farmers. The chatbot suggested herbicides but got the timing of application wrong, which the article says could cause significant crop damage. Researchers also found inaccurate information on planting time, seed rate, and fertiliser rate and timing.
• A canning instruction that makes no sense. The same article quotes an extension educator who asked two chatbots about home canning and was told to put tomatoes in an ice bath. He called it a silly reply to a serious question. Canning errors carry a food-safety risk.
• Popular does not mean effective. He also warned that a chatbot choosing the most commonly mentioned pesticide from the web may not be picking an effective one.
• Hedged and generic answers. Norm, a farm-data chatbot from the agricultural data platform Farmers Business Network, gave a vague reply on how often to spray corn herbicide, saying frequency "can depend on several factors", and its interface carries its own caution that it is "not intended for real agronomic guidance".
• Plain falsehoods on contested topics. Ambrook's October 2023 piece quotes a farmer who says that on animal agriculture, sustainability and GMOs, ChatGPT produces things that are "blatantly false" from his standpoint. The head of data science at Farmers Business Network is quoted warning against a chatbot "hallucinating recommendations for highly regulated products". That was a stated risk, not a documented incident.

Two honest limits. These reports use earlier versions of the tools, and I could not find recent independent testing across farming topics, so I cannot tell you today's error rate. And several of the examples come from advice to farmers in other growing conditions than yours, which is itself the problem: a general chatbot does not know your soil, variety, weather or local product approvals unless you tell it, and even then it may not apply them correctly.

Why it errs on exactly these questions

Look at what the failures have in common. Each depends on something local and specific: the growth stage of this crop on this field, the approved products for this use, the label for this formulation. The chatbot has seen text about herbicides in general and fills in a typical answer. Typical answers are right often enough to be trusted and wrong often enough to hurt.

Three traps follow from that.

1. Averaging. The "usual" timing or rate across many sources may not be correct for your variety, region or season.
2. Staleness. Approvals, withdrawal periods and rules change. A chatbot trained on older text, or on a mix of dates, does not flag that.
3. Misplaced certainty. Ask for a rate and it gives a number. It rarely says "I don't know" unless you force the question.

Safe jobs and unsafe jobs, sorted

The sorting rule is whether a person reads and checks the output before anything is done with it, and whether a mistake can be undone.

The pattern in the safe rows: words, spreadsheets and forms. In the unsafe rows: numbers that act on living things or the food chain.

A worked example: the same question, handled two ways

An illustration. A mixed arable farm has a patch of grass weeds in a winter wheat field. The farm manager is tempted to ask ChatGPT what to spray and how much.

The risky way. The manager types: "What should I spray on blackgrass in winter wheat and at what rate?" The reply names a product class, a rate per hectare, and a "best time" in early spring. It reads well. If the manager follows it, there are four ways to be wrong: the product may not be approved for that crop stage, the rate may differ from the label, resistance on this farm may make the product useless, and the timing may miss the weed's growth stage.

The safer way. The manager asks for help with the conversation, not the decision:

I farm winter wheat and have a patch of grass weeds I think is blackgrass, in one field. I am going to see my agronomist.
Write me a list of questions to ask, covering: how to confirm the species, whether herbicide resistance is likely on my farm and how to test for it, what non-chemical options exist, what the label and local approvals will restrict, and what records I must keep afterwards.
Do not recommend products or rates.

A realistic (illustrative) reply is a 12-question list: how to confirm identification, whether a resistance test is worth doing, what the current growth stage allows, whether rotation or a delayed drilling date would help, what the buffer and application restrictions are on the label, and what to record afterwards. What you fix: remove any question that assumes a particular product, and add your own field history. Notice the difference: the assistant has made the manager better prepared for the adviser, and has not replaced them.

The five-question test to run on your own farm

Before you trust ChatGPT with any farm task, ask it five questions to which you already know the answer from your own records and your adviser. Do it in an evening, write down the results, and repeat when a new version is released.

1. A fact from your own farm. Ask for something you told it earlier or that you know, such as your last wheat variety. This tests whether it invents details when it does not know. A good answer says it does not have the information.
2. A rate you can read off a label. Pick a product you use and ask for the label rate on a named crop. Compare with the label. If it is wrong, or it gives a figure without saying to check the label, you have your answer for every future rate question.
3. A timing you know well. Ask for the latest safe date to apply something or turn stock out, using a case where you know the correct answer for your area and season. Check whether it asks for your location and stage or just answers.
4. A trap question. Ask about a product or disease that does not exist, such as a made-up product name. A reliable tool says it cannot find it. An unreliable one writes a confident description.
5. A "say when you don't know" test. Ask a question that has no single answer, such as what the best fertiliser plan is for "a field" without any details. Does it ask for soil results, crop, previous crop and region, or does it hand over a plan?

Score yourself: a tool that fails any of questions 2, 4 or 5 should not be used for anything that acts on crops or animals. A tool that passes all five has earned the right to help with drafting and explanation, not the right to decide. And passing today is not a guarantee about next month: models change, so rerun the test.

Safe uses worth setting up properly

Farm admin is where an assistant earns its keep, and the commonly suggested uses in the farming press are paperwork for commodity groups and government schemes, and marketing copy. Here are four with a short example each.

Grant or loan answers from your own facts

A farmer gives the assistant the farm's acreage, enterprises, last three years' turnover and the planned investment, and asks for a first draft of the "how will this improve the business" answer. The draft is only as true as the facts, so the farmer checks every figure against the accounts. The wider job of drafting a plan for a lender is in AI farm business plan: draft one for a lender from real records.

Rotas, job messages and harvest-time communication

Seasonal staff need clear start times, tasks and kit lists. The assistant turns your bullet notes into a short message for each group, and rewrites it in simpler words when some staff have English as a second language. Sending is yours. See AI automation for agriculture: rotas, timesheets and harvest messages for the set-up.

Tidying records you already keep

Milk, feed or yield records that live across several spreadsheets can be cleaned and summarised. Ask for totals and a table, then verify the totals yourself with a sum. AI tools for dairy farming: analyse milk and DHI records in ChatGPT gives a worked version with the checks.

Plain-language help with paperwork terms

A label or scheme document written in dense legal language can be rewritten in short sentences. This is a reasonable use because you are asking for explanation of a text you give it, and you can compare the two. The rule: the original text is what counts, and if the explanation and the document disagree, the document wins.

A filled-in prompt for the safe side: a supplier query

To show what a good safe-use prompt looks like, here is one for an ordinary admin job, with the guard rails written in.

Draft an email to our seed merchant. Facts: we ordered 12 tonnes of winter barley seed on [date], order ref [ref], promised delivery [date], now [number] days late; drilling is planned for [date]. We want a confirmed delivery date within 48 hours, and a part-delivery if possible.
Tone: firm and polite. Do not mention compensation. Do not state any seed rate, variety trait or treatment. If a fact is missing, write [CHECK] instead of guessing.

A realistic (illustrative) output is a 90-word email that quotes the order reference, states the lateness and the drilling date, and asks for the two things. What you fix: any sentence that sneaks in a claim, such as "as this seed is vital for a good yield". The assistant had no basis for it, and a supplier may read it as an opening to argue.

A quick sum on admin time

An illustration with invented figures. A 300-hectare mixed farm spends about 6 hours a week on emails, scheme forms, rota messages and record tidying, usually in the evening. If an assistant halves the drafting part of that, say 3 hours of the 6 become 1.5, the saving is 1.5 hours a week, about 75 hours a year. Valued at even a modest $20 an hour, that is $1,500 of time, and it comes back as evenings rather than as cash. Set that beside the cost of one wrong spray decision and the argument for using the tool where it is safe, and not where it is not, makes itself.

When a farm app beats ChatGPT

A general chatbot knows a little about everything. A farm app knows your holding. If the job relies on your own live data or on advice tested for your conditions, an app is the better tool.

• Livestock records and treatments. Herd or flock apps keep animal IDs, treatments, withdrawal dates and movements in one place and can flag when a withdrawal period has not ended. A chatbot has no memory of your herd unless you paste it in every time, and no duty to be right about a date. A comparison is in Herdwatch, AgriWebb and Farmonaut compared.
• Field monitoring. Satellite and machine-data tools see your fields. A chatbot sees only what you describe.
• Agronomy built on trials. Tools grounded in checked data are built to cut misinformation. Norm, for example, is described as drawing on public weather, soil and product-label data plus its maker's own agronomic data and government statistics, yet it still tells users it is not for real agronomic guidance. Grounded is better, not infallible.
• Anything with a legal record. Treatment books, movement records, spray records: use a tool built to store and export them.

A decision rule: if the question is "what does the law, the label or my own record say?", use the source or a tool wired to it. If the question is "help me write, sort or explain this", ChatGPT is fine. For a comparison of assistants and farm apps side by side, see AI assistant for farmers: ChatGPT, Claude or a farm app?, and for what a small farm gets without paying, free AI for agriculture.

How to ask so the answer is as safe as it can be

Even for the safe jobs, a few habits cut the risk of a confident error getting through.

• Give facts, ask for structure. Supply the figures and ask the assistant to arrange them, instead of asking it to supply the figures.
• Tell it what to do when it does not know. Add: "If you are not sure, say so and list what you would need to know. Do not guess." It does not make the tool omniscient, but it shifts the default.
• Ask for the source. For any factual claim, ask where to check it. If the answer is vague, treat the claim as unverified.
• Name your location and season so it does not answer for somewhere else, and still check the result, since naming a place does not guarantee local accuracy.
• Keep a note of what you accepted. If you used a figure from an assistant for anything, write down which and where you verified it. It matters if a record is ever questioned.

A mistake that came from a sensible-looking summary

As an illustration of how this goes in practice, a small livestock farm asks an assistant to summarise a vet's written treatment plan in simple words for the farm staff. The summary reads well and is pinned in the shed. It leaves out one line: a withdrawal period after the last dose before milk or meat can enter the food chain. Nobody notices, because the summary is shorter and cleaner than the original. The farm catches it only when the vet's follow-up visit compares the sheet with the plan.

The repair is cheap: any summary of a treatment instruction is checked line by line against the original by the person who wrote it, and the original stays attached. The principle is that summarising is not harmless when a missing line is the safety line. Where summaries touch medicines, chemicals or legal records, give the assistant the job of formatting, and give the vet or adviser the job of confirming.

What to say to staff, in a paragraph

If several people on the farm use ChatGPT, agree a short rule and write it on a card: "Use it for words and spreadsheets. Never use it for doses, rates, timings or diagnoses. If you used it for anything that goes out to a customer, a lender or an official body, someone else reads it first." It is easy to remember and easy to apply, and it keeps the benefit of the tool while closing the dangerous door. Revisit it each time you add a new use, and rerun the five-question test when a new version arrives.

Questions farmers ask about ChatGPT

Can I ask ChatGPT how much pesticide or medicine to use?

Do not take doses, spray rates or withdrawal periods from a chatbot. Those figures are fixed by the product label and your adviser or vet for your crop, animal and conditions. A chatbot can explain what a label term means or help you list questions to ask, but the number comes from the label.

Is ChatGPT better now than in the studies?

Likely better in places, but I could not find recent independent tests across farming topics, so the safe assumption is that it still makes confident errors. The rule that matters does not depend on the version: anything where a wrong answer costs a crop, an animal or a licence is checked against the label or a qualified adviser.

What can ChatGPT safely do on a farm?

It is safe for jobs where you read the result before use: drafting emails, grant and loan answers, job descriptions, rotas, supplier chasers, social posts, and tidying records you supply. Commodity-group and government paperwork drafts are a commonly suggested use, provided you check every figure and date.

When is a farm app better than ChatGPT?

When the job needs your own live data or a vetted source: livestock records, field maps, satellite imagery, machine data or agronomy built on tested local trials. A farm app draws from your records and its own checked content, while a general chatbot answers from patterns in its training text.

Further reads

• Can Farmers Use AI for Farm Records and Paperwork? — Use AI for farm records and paperwork, with checks.
• AI Chatbot for Farmers: Answers From Your Own Farm Notes and SOPs — A chatbot that answers only from your own notes.
• Livestock Record Keeping App: Herdwatch vs AgriWebb AI Voice Notes — Livestock records by voice, in a purpose-built app.
• Farmonaut Pricing vs FieldView Plus: What Satellite Monitoring Costs — What satellite field monitoring costs.
• Farm Accounting Software: QuickBooks, Xero or a Farm App With AI? — Where the money side of the farm should live.
• AI for Hydroponic Farming: Review Nutrient and pH Logs Each Week — A grower's daily log becomes a weekly exception review, with readings checked against the crop's own target sheet.
• AI Chatbot for Farming: What It Can Answer and Where It Fails — What a chatbot can reliably do on a farm, with sample prompts, and the agronomy, livestock and safety questions it should never settle.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Offrange (formerly Ambrook Research), 'For Farmers, Are AI Chatbots Worth the Risk?' (9 August 2024) and 'Farmers Are Tentatively Embracing AI' (6 October 2023); a Nature Food study of ChatGPT advice for cassava farmers as reported there.