Lesson objective

Recognise the specific ways a generative tool invents products, rates, restrictions and rules in green-industry work, and apply a verification step that catches each one.

Plain-language explanation

Why this matters now

Ask a general assistant about a weed in cool-season turf and you will get an answer immediately. It will be in complete sentences, it will use the right vocabulary, and it may name a product, a rate, a timing window and a site restriction. Nothing about the way it arrives signals which parts of it are real.

That is not a bug someone will fix next year. It follows from how these systems are built, and the federal standards body has written it down.

The consequence for a landscape, lawn care or tree care business is sharper than for an office business, because several of the things a model is most likely to invent are things that carry legal weight: a product name, a registration, a rate, a site restriction, a re-entry interval, a licence requirement.

A question to sit with

If an assistant gave you a product name, a rate and a restriction, and one of those three was invented, which one would cost you the most, and how would you have found out?

What goes wrong

Invented products. A model can produce a plausible trade name, sometimes by combining parts of real ones. It reads like a product because it is built from text about products.

Invented registrations. An EPA registration number has a recognisable shape. A model can produce something of that shape that belongs to nothing. A number that looks right is not a number that is right.

Invented rates and intervals. Rates, dilution ratios, re-entry intervals and pre-harvest or pre-grazing intervals are numbers embedded in text. They are exactly the sort of thing prediction produces fluently and wrongly.

Invented rules and bodies. A model may state that a state licenses something it does not license, that a body has issued AI guidance it has not issued, or that a standard has a clause it does not have. None of the green-industry regulators or bodies has published AI guidance at all, so any output claiming such guidance is wrong before you check anything else.

Invented sources for the invention. Because outputs may include "confabulated logic or citations", asking the model where it got the answer can produce a second invention that makes the first look verified.

Smoothed-over gaps. Asked a question it cannot answer from what you supplied, a model will often answer anyway rather than say the material does not contain it.

The practical answer

Sort the content of any output into three buckets before you use a word of it.

Bucket Examples in this trade What you do
Must be verified against a primary document Product name, registration number, rate, dilution, site restriction, re-entry interval, licence requirement, contract term, price Read it on the label, the specification, the state agency's own page, or your own price book. An AI output is never proof
Must be recomputed Areas, volumes, per-thousand quantities, crew hours, totals on a proposal Recompute outside the tool, in your estimating software or on paper
Judgement you own Whether a plant is worth saving, whether a tree is safe, whether to treat at all, what to promise a customer Made by the qualified person, not delegated to an output

Nothing in the first two buckets is ever finished by the model. That is the whole method.

The five-question check, for any output that touches a living thing or a rule.

  1. What in this is a name, a number or a rule? Underline each one. Those are your verification targets.
  2. Where would a person prove this? Name the document: the container label, the state pesticide agency's own listing, the manufacturer's published label, the property specification, your price book.
  3. Did I supply that document, or is the model remembering? If nobody supplied it, treat every fact as unverified, not as probably right.
  4. What did it not say? Look for the missing caveat, the missing condition, the missing "it depends on the site".
  5. Who owns the decision this output leads to? For a pesticide, the certified applicator, or the certified applicator supervising a non-certified one, owns the product, rate, timing, site and conditions, and the product label is the binding instruction: using a registered pesticide in a manner inconsistent with its labeling is an unlawful act under federal law. For a tree's condition, the arborist, or in a state that licenses tree work the licensed tree expert. For a promise, the person who signs.

Asking a question that is answerable. You cannot stop a model confabulating, but you can make its answer checkable. Three patterns help.

Instead of Ask Why it helps
"What rate should I use on this weed?" "Using only the label text pasted below, list every statement about rate on cool-season turf and quote each one word for word. If it is not stated, reply 'not stated'." Gives it a source you hold and a way to decline
"Does my state require a licence for this?" "List the questions I should ask my state's pesticide agency to find out whether this work requires certification, and what to have ready when I call." Turns a fact you cannot verify into a task you can do
"Summarise this specification." "Summarise only what the pasted specification says. Mark with [NOT IN DOCUMENT] anything a reader might expect that the document does not cover." Surfaces the gap instead of smoothing it

Module 6 builds these into a full prompt structure. The point here is narrower: a question that can only be answered from material you hold is a question whose answer you can check.

The frontier is uneven, and you cannot see it. A preregistered experiment offers a useful shape, by analogy only. Dell'Acqua and colleagues, in Navigating the Jagged Technological Frontier, published in Organization Science, volume 37, issue 2 (March 2026, day not stated; earlier Harvard Business School Working Paper 24-013 of 22 September 2023), randomly assigned 758 consultants at Boston Consulting Group to no AI, GPT-4, or GPT-4 with a prompt engineering overview. Inside the AI capability frontier, across 18 tasks, AI users completed 12.2% more tasks 25.1% more quickly with significantly improved quality. On a complex task deliberately chosen to sit outside the frontier, subjects using AI "were 19% less likely to produce correct solutions compared with those without AI". Those were consultants on business tasks with GPT-4 in 2023, not landscapers on a property, and the analogy is all that carries: help and harm sit next to each other, the boundary is not marked, and the output looks the same on both sides.

Remember this

  • Confabulation is a natural result of how generative models are designed, not a rare fault; NIST publishes no rate for it and neither should you.
  • Products, registrations, rates, intervals, licence rules and cited sources are all things a model can invent fluently.
  • Sort every output into verify, recompute and own, and finish the first two outside the tool.
  • Asking the model where it got an answer can produce a second invention that dresses up the first.
  • No green-industry regulator or body has published AI guidance, so an output claiming such guidance is wrong on its face.

Landscaping example

Thornhedge Grounds Care is a fictional maintenance company. Account manager Luz Ferreira is writing a renewal proposal for the invented Ambergate HOA and wants a paragraph explaining a broadleaf weed problem in the common areas. She asks the company's approved assistant for an explanation and a recommended treatment. The reply is three tidy paragraphs. It names a product, gives a registration number in the familiar format, states a rate per 1,000 square feet, gives a re-entry interval in hours, and adds that "the state requires a 24-hour posting notice for all residential common-area applications". Luz applies the five-question check. Underlining the names, numbers and rules gives her six items to verify. She asks where a person would prove each: the label for the product, rate and re-entry interval, the state pesticide agency for the posting rule, and Thornhedge's own price book for the cost. She then notices the answer to question three: she supplied nothing. No label, no specification, no agency page. Everything in the reply is the model remembering. She takes it to Thornhedge's certified applicator, Ines Okonkwo, who reads the product name and says she has never heard of it. Ines checks the company's inventory and the manufacturer labels she holds, and the product is not among them; nothing with that name appears on the state agency listing she uses either. The registration number has the right shape and belongs to nothing she can find. Ines also says the posting claim is not something Luz should ever take from an assistant, because no green-industry regulator has issued AI guidance and the rule, if it exists, comes from the agency and the label. Luz rewrites the paragraph with no product, no rate and no legal claim in it: a plain description of the weed pressure, the site conditions, and a line saying a certified applicator will select any product and set any rate from the label. Ines signs off on the treatment approach separately, from the labels in her hand.

Responsible-use warning

Do not treat a product name, a registration number, a rate, a re-entry interval or a licence rule from an AI output as a starting point that only needs a quick sanity check. It is not a draft of a fact; it may be nothing at all, wearing the right shape. Ask where a person would prove it, and prove it there: the label on the container, the manufacturer's published label, your state agency's own information, the property specification, your own price book. Asking the model for its source can produce a second invention. And never put a rate or a legal claim into a customer document at all if a qualified person has not verified it.

Human-review requirement

Before any AI-assisted text reaches a customer, a record or a crew, a named qualified person verifies every name, number and rule in it against a primary document, and the verification is recorded with the document and the date. Where the output touches a pesticide, the certified applicator, or the certified applicator supervising a non-certified one, owns the product, rate, timing, site and conditions, and the label on the container is the binding instruction. Where it touches a tree's condition, the arborist or the licensed tree expert owns it. Where it makes a promise or a price, the person who signs owns it. The reviewer confirms nothing was accepted because it sounded right.

Practical exercise

Run a confabulation hunt on material you control. Use an invented property and an invented specification only, and do not put any customer detail into a tool for this exercise.

  1. Write a short fictional site description of about 100 words: an invented property, a turf type, a weed or pest pressure, a season and one site constraint such as a pond edge or a school playground.
  2. Ask your approved assistant to explain the problem and recommend a treatment. Save the reply exactly as it came.
  3. Underline every product name, number, interval, percentage, licence claim and legal claim in the reply. Count them.
  4. For each underlined item, write the document a person would use to prove it, and who in your company would hold that document.
  5. Mark each item as supplied by you, or remembered by the model. Note how many items fall in the second group.
  6. Ask the assistant for its sources for three of the items. Record what comes back and write one sentence on whether that made anything more verifiable.
  7. Ask the same question again with the pasted specification and an instruction to reply "not stated" where the document is silent. Compare the two replies.
  8. Write the paragraph you would actually put in front of a customer: no product, no rate, no legal claim, and a named qualified person taking the decision.
  9. Record what you verified, against what, on what date, and by whom, using the identification verification sheet or the workbook.

Knowledge check

  1. How does NIST AI 600-1 define confabulation, and what does it not tell you?

    Answer guidance: It defines it as the production of confidently stated but erroneous or false content, known colloquially as hallucinations or fabrications, and says it is a natural result of the way generative models are designed. It does not quantify how often confabulation occurs, so no hallucination rate may be attributed to it.

  2. Why is asking a model for its sources a weak check?

    Answer guidance: Because outputs may include confabulated logic or citations that purport to justify or explain the answer. A second invented source can make the first invention look verified. Verification means opening the primary document yourself.

  3. An output states that a green-industry body has issued guidance on using AI in plant health care. What is wrong with it?

    Answer guidance: No green-industry regulator or professional body has published AI guidance. The obligations in this trade are the ordinary application of label law, licence scope, supervision duty, safety duty, employment law and deception law, so the output is wrong on its face.

  4. What does the jagged frontier experiment suggest about where AI helps, and how far does it transfer?

    Answer guidance: With 758 Boston Consulting Group consultants, inside the frontier AI users completed 12.2% more tasks 25.1% faster with better quality, while on a task outside the frontier they were 19% less likely to produce correct solutions. They were consultants using GPT-4 in 2023, so it transfers only as an analogy: the boundary is unmarked and the output looks the same on both sides.

Sources and further reading