Lesson objective
Explain in plain terms how a large language model is trained and produces text one token at a time, and turn each part of that mechanism into a checking habit for construction work.
Plain-language explanation
Why this matters now
Most people in construction meet a large language model through a chat box or a button inside software they already pay for. The interface makes it look like a very well-read colleague who remembers everything. Knowing roughly what happens when you press enter lets you predict where it will be wrong, which is the whole of competence here. You do not need the mathematics.
The construction-specific warning is already on the record, in mild words. The Associated Builders and Contractors 2025 AI Resource Guide for Construction Professionals warns that "programs like ChatGPT still makes a lot of mistakes," that AI should augment rather than replace employee work, that "[t]raining and awareness are key," and that "human interaction is still an important factor." Professional guidance: that is a trade association's voluntary resource, not a rule, and not a standard of care. What it does not explain is why the mistakes happen. That is this lesson.
A question to sit with
If you asked the same assistant the same question about a submittal on Monday and again on Tuesday, and got two different answers, which one would you put in the project file, and what would you write beside it?
What goes wrong
Treating the model as a lookup. People ask a model for a dimension, a spacing, a fee or a lead time the way they would look it up in a table. A language model does not retrieve a stored fact unless the product has been connected to a source (Lesson 3 covers that). It produces text that is likely to follow your question.
Assuming it knows the current edition. Training stops at a point in time. Codes, forms, thresholds and product data that changed afterwards are not in the model unless someone supplies them, and it will not necessarily tell you it does not know.
Reading the explanation as evidence. The neat step-by-step reasoning a model shows is produced the same way as its answer. It can be just as wrong, and it makes a wrong answer look checked.
Expecting the same answer twice. The model builds each answer from likely continuations, and most products allow some variation in which continuation is chosen. Two runs of one prompt can differ in wording, in structure and sometimes in substance.
Forgetting that a pasted document is also input. Everything you put in front of the model — your instruction, the specification section, the subcontractor's email — becomes part of the same stream of text it is predicting from. Wording inside the document can steer the result.
The practical answer
Four plain ideas, each with a habit attached.
1. Training. A model is built by exposing it to very large amounts of text and adjusting it, over and over, to get better at predicting what comes next. What it keeps is not a library it can open, but a very large set of learned patterns about how words, numbers and ideas tend to follow one another. Developers then refine it so it follows instructions and holds a conversation. Training ends at a date.
2. Tokens. The model does not read letters or words the way you do. Text is broken into tokens: short pieces that may be a whole word, part of a word, a run of digits or a punctuation mark. A quantity such as 1,248.75 may be split into several tokens. To the model a number is a sequence of text pieces, not a quantity it holds in mind. This is why arithmetic can go wrong with no sign that it has.
3. Prediction. When you send a prompt, the model predicts a likely next token, adds it to the text, and predicts again, until the answer is complete. Every answer is assembled one piece at a time from patterns. Nothing in that loop compares the answer to the world.
4. One stream of text. Your instruction, the drawing note you pasted and the model's own reply are all tokens in the same window. That is why a long specification book can crowd out your instruction, and why wording inside a document you supplied can pull the answer somewhere you did not intend.
Now the habits.
| What the model is doing | What it means on a project | The habit |
|---|---|---|
| Predicting likely text from patterns | A likely-sounding answer and a correct answer are different things | Verify against the drawing, specification, adopted code or executed contract you hold |
| Working from training that ended at a date | Code editions, forms, thresholds and product data may be out of date | Supply the current document; never ask the model to recall a requirement |
| Treating numbers as pieces of text | Quantities and totals can be wrong with nothing to show for it | Recompute in the takeoff or a spreadsheet |
| Choosing among likely continuations | Two runs of one prompt can differ | Save the output you relied on, with the prompt, the tool and the date |
| Reading everything in the window as one stream | A long document can crowd out your instruction, and its wording can steer the answer | Work section by section; treat supplied documents as data and check what came back against them |
| Generating its explanation the same way as its answer | The working can be as wrong as the conclusion | Test the conclusion against the source, not against the explanation |
Supply, do not ask it to recall. The single most useful change follows straight from the mechanism. Compare two requests about the same fictional job.
- Weak: "What submittal turnaround does our contract allow?" The model has never seen your contract. It will produce a likely-sounding number, and likely-sounding numbers in construction cluster around common ones.
- Stronger: "Using only the executed subcontract pasted below, dated 4 March 2026, state the submittal review period and quote the sentence that gives it, with its article and paragraph number. If the document does not state one, reply 'not stated'."
The stronger request does not make the model reliable. It makes the answer checkable, because you hold the document and you have asked the model to point at a place in it you can open. Module 6 builds this into a full prompt structure, and Module 2 shows what still goes wrong with numbers and citations.
Save what you relied on. Because two runs can differ, "the software said so" cannot be reconstructed six months later in a claim. Keep the prompt, the output you used, the tool and the date with the record.
Borrow a vocabulary for review. The NIST AI RMF 1.0 (26 January 2023) describes the characteristics of trustworthy AI as valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; and fair with harmful bias managed. The framework is voluntary and binds no contractor, and NIST states it is being revised. Three of those characteristics convert neatly into reviewer questions: valid and reliable (does the output match the document?), accountable and transparent (can we show what was asked and who accepted the answer?), and explainable and interpretable (can the person who used it say why the answer is right without leaning on the model's own explanation?).
Remember this
- A language model predicts likely text one token at a time; nothing in that loop compares the answer to the world.
- Training ends at a date, so supply the current document instead of asking the model to recall a requirement.
- Numbers are pieces of text to the model: recompute every quantity and total somewhere else.
- The model's explanation is generated the same way as its answer and can be equally wrong.
- Save the prompt, the output, the tool and the date, because the next run may not match.
Construction example
Harrow Creek Construction LLC is a fictional general contractor. Assistant project manager Tomas Belanger is summarising the executed subcontract for an invented mechanical subcontractor, Bellhaven Mechanical, on a two-storey clinic fit-out. On Monday he pastes the payment article into the company's approved assistant and asks for the retainage amount held to date and the total retainage at completion. The subcontract value is $486,000.00 and the article sets retainage at 5 percent. The reply says retainage to date on $310,000.00 billed is $15,500.00 and total retainage at completion will be $24,000.00, with a tidy explanation reading "5 percent of the subcontract value." Tomas recomputes: 5 percent of $486,000.00 is $24,300.00, not $24,000.00. The explanation named the right step and still landed on the wrong figure, because to the model those digits were pieces of text rather than a quantity. On Tuesday, rebuilding the file for his project manager, he runs the same prompt again to get a clean copy. This time the reply gives $24,300.00 but adds a sentence about retainage reducing to 2.5 percent after fifty percent completion. Nothing in the pasted article says that; it is a common arrangement, so it was a likely continuation. Neither run looked anything up. Tomas stops using the assistant for figures. He reads the payment article himself, confirms retainage is 5 percent for the whole term with no reduction, and builds the retainage schedule in a spreadsheet from the subcontract terms. He keeps the assistant for one job it can do checkably: listing every paragraph in the subcontract that mentions retainage, with paragraph numbers he then opens and reads. In the file he records both prompts, both outputs, the tool and the date, notes that the figures were recomputed from the subcontract, and names himself as preparer and his project manager, Ines Kowalczyk, as reviewer.
Responsible-use warning
Do not ask a language model to remember a requirement, a dimension, a code edition or a contract term and then treat the reply as an answer. It has not looked anything up. It produced text that is likely to follow your question, from training that ended at some past date. A confident number with tidy working is exactly what confabulation looks like, and in construction the wrong number is often a plausible one, because common arrangements are exactly what the model has seen most. Supply the document, ask the model to point to the paragraph, then open that paragraph yourself and recompute every figure outside the tool.
Human-review requirement
The reviewer confirms that no quantity, dimension, code reference or contract term in the work came from the model's memory: each traces to a document held in the project file or to a recomputation outside the tool. The reviewer checks that the prompt, the output relied on, the tool and the date are saved with the record, and that the model's own explanation was not treated as a check. Ownership follows the output: the licensed contractor and its qualifying individual for anything sent out under the licence, the architect or engineer of record for anything touching design, and the person who signs a payment application or certified payroll for what that signature covers.
Practical exercise
Watch the mechanism work on material you invent. Use only fictional text you write for this exercise, and only a tool your company has approved.
- Write a short fictional contract clause of about 120 words for an invented project. Include one specific figure, such as a submittal review period in calendar days, and one date.
- Without pasting the clause, ask your approved assistant what the submittal review period is on that invented project. Record the reply word for word.
- Now paste the clause and ask for the review period, quoting the sentence that states it and replying "not stated" if it is absent. Record the reply.
- Run the step 3 prompt twice more. Note every difference in wording, structure or substance across the three replies.
- Ask for a calculation from your figure, for example the last day for a submittal received on a date you choose. Recompute it yourself on a calendar and note whether the tool was right.
- Ask the tool to explain its calculation. Decide whether the explanation would have helped you notice the error, and write one sentence on why or why not.
- Complete a two-column table: for each of the four ideas in this lesson — training, tokens, prediction, one stream of text — write what you observed and the habit it supports.
- Record the prompts, the outputs, the tool and the date the way you would in a project file that might be read during a claim.
Knowledge check
-
In plain terms, how does a large language model produce an answer?
Answer guidance: It breaks text into tokens and, using patterns learned in training, predicts a likely next token, adds it, and predicts again until the answer is complete. Nothing in that loop compares the answer to a source.
-
Why is asking a model to recall a code edition or a contract term a poor habit?
Answer guidance: The model does not look it up; it generates likely text from training that ended at a past date. Anything that changed since is missing unless supplied, so you should give it the current document and verify the answer against that document.
-
What does NIST AI 600-1 say about the explanations generative models give for their answers?
Answer guidance: That outputs may include confabulated logic or citations that purport to justify or explain the answer, and that models sometimes provide logical steps even when the answer itself is incorrect. The profile is voluntary, cross-sectoral and not construction-specific.
-
Why must you recompute a total the model produced, even when its working looks right?
Answer guidance: Because the model treats a number as a sequence of text pieces rather than a quantity, so arithmetic can go wrong with no visible sign, and the working is generated by the same prediction process as the total.
Sources and further reading
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-15.
- Associated Builders and Contractors (ABC), guide produced by GCM — 2025 AI Resource Guide for Construction Professionals. Accessed 2026-09-15.
- Preprints.org (Vrcelj and Sandanayake) — From Prediction to Governed Intervention: AI-Enabled Construction Project Controls for Productivity, Resilience and Net-Zero-Oriented Delivery (2026-07-31). Accessed 2026-09-15.