A carbon accountant has a supplier's emissions response open in one window and an AI chat in the other. The supplier wrote, in a PDF, that their site's Scope 1 emissions for 2024 were 4,210 tonnes CO2e. The accountant, busy, types into the chat: "what were this supplier's 2024 Scope 1 emissions?" The model, which has never seen the PDF, answers "approximately 5,000 tonnes, typical for a manufacturer of this size." It is fluent, it is plausible, and it is pure invention. The right number was sitting in the document twelve inches away. The model answered from its imagination because nobody forced it to the page. This lesson is about that forcing: how to make the model read THE document instead of recalling a world that does not include your supplier.

Two Ways a Model Can Answer: Grounding Versus Recall

Every answer a language model gives comes from one of two places, and the difference between them is the difference between a defensible disclosure and a liability. The first is recall: the model produces an answer from the patterns baked into its training, a kind of compressed average of the open web and everything else it read. The second is grounding: the model produces an answer from specific source text you have placed in front of it in the conversation, reading and quoting that text rather than its memory.

For a disclosure professional, this distinction is not academic. It is the whole game. A figure produced by recall is, by construction, untraceable: it came from a blend of sources you cannot name, for a company that is not yours, in a year that may not be your reporting year. A figure produced by grounding came from the supplier response, the utility bill, the policy, or the prior report that you handed the model, which means it has a source you can point to, quote, and put in the assurance file. The entire technique of this lesson is one sentence long: stop the model from answering by recall, and force it to answer by grounding.

Why does this matter so much more for disclosure than for, say, drafting an email? Because in disclosure the value of an answer is almost entirely a function of where it came from. A correct-sounding emissions figure with no source is worthless to you; a figure that traces to page 3 of the supplier's signed response is gold, even before you have checked it, because it is the kind of thing that can be checked. Grounding does not just make answers more accurate. It makes them the right species of thing: evidence, rather than opinion.

Why Open-Web Recall Is Poison for a Disclosure Number

It is worth being blunt about why recall, the very thing that makes a chatbot feel magical, is poison for a reporting figure. When you ask an ungrounded model for a number, three failures are baked in at once.

First, it is not your data. The model has never seen your supplier's actual 2024 emissions. It produces a number representative of "a manufacturer like this," which is precisely an industry average dressed as a specific fact. In Scope 3 work, where the temptation to let AI fill the gaps with averages is already the field's most dangerous habit, an ungrounded answer is that exact temptation arriving uninvited.

Second, it has no date and no version. Emission factors are revised; databases publish annual editions; policies are superseded. A recalled figure is a smear across whatever years the model trained on, with no way to know if it matches your reporting period. The same prompt next month might return a slightly different number, and you would have no way to explain the change.

Third, and most dangerously, it is untraceable by design. There is no page, no row, no document to point an assurer at, because the answer was assembled from a statistical blend, not read from a source. This is the property that ends careers: a number that cannot be traced is, in an assured disclosure, indistinguishable from a number that is wrong, because you cannot prove it is right. The assurer does not need to show your figure is false. They only need to find that you cannot show it is true.

An ungrounded answer is an industry average wearing the costume of a specific fact. In disclosure, force the model to the document, because a number you cannot point at is a number you cannot defend.

The cure is not a better model or a cleverer question. It is to remove recall from the equation entirely by putting the real document in front of the model and confining it to that text. The rest of this lesson is the mechanics of doing exactly that.

The Paste-the-Source Technique

The most reliable way to ground a model on a document is also the most direct: paste the relevant source text into the conversation, then instruct the model to answer only from that text. There is no magic to it, and that plainness is the point. You are converting the model from a thing that remembers into a thing that reads.

The structure has three parts, and each one is load-bearing.

Paste the Actual Text

Put the source material directly into the prompt: the paragraph from the supplier's response, the line items from the utility bill, the relevant clause of the policy, the figure and footnote from the prior report. Give the model the real words, not a description of them. "The supplier said their emissions were about 4,000 something" reintroduces your memory as an error source; pasting the exact line removes it. If the document is long, paste only the relevant section, clearly delimited, so the model is not tempted to wander.

Answer Only From the Text Above

Then constrain the model with an explicit instruction: answer only from the text above; if the answer is not in the text, say "not stated in the document" and do not supply a figure from elsewhere. This sentence is the wall between grounding and recall. Without it, a model handed a document will often blend what it read with what it remembers, producing an answer that is partly grounded and partly invented, which is the worst of both worlds because you cannot tell which part is which. The instruction forces a clean line: either the answer is in the provided text, or the model declines.

Quote the Exact Line

Finally, demand that the model quote the exact line or sentence it relied on, verbatim, before giving its answer. This is the technique's quiet masterstroke. A quote is checkable in a way an assertion is not. When the model says "the document states: 'Site Scope 1 emissions for 2024 were 4,210 tCO2e' (page 3), so the figure is 4,210 tCO2e," you can do something no other output lets you do this cheaply: glance at page 3 and confirm the quote is really there and really says that. The quote also catches the subtler failure where a model paraphrases a document into something it does not quite say, because the verbatim text either supports the answer or visibly does not. You are not trusting the model's summary. You are trusting your own eyes on a quote the model has helpfully located for you.

The Documents You Force the Model To

The technique is general, but it helps to see it land on the specific documents a discloser lives in, because each carries its own trap that grounding defuses.

The supplier response is the heart of Scope 3 and the place an ungrounded model does the most damage, because it will happily invent a "typical" figure for a supplier it has never heard of. Grounded on the actual response, it extracts what the supplier actually reported, and the quote preserves whether that figure was supplier-measured or supplier-estimated, a distinction your file depends on.

The utility bill is your Scope 2 activity data at its source. Pasted in, the model can pull the exact kWh and the billing period; recalling, it would guess a consumption "normal for a building like this," which is no use at all. The exact line from the bill is the activity data; anything else is fiction.

The policy, whether your own or a regulation, is where paraphrase is most dangerous, because a softened or sharpened restatement of a requirement can quietly change what you think you must disclose. Forcing the model to quote the clause verbatim before interpreting it keeps its interpretation tethered to the actual words.

The prior report is your reconciliation anchor and your consistency check. Grounded on last year's actual disclosure, the model can tell you what you said before; ungrounded, it would invent a plausible prior figure, which is worse than useless because it looks like a reconciliation while being a fabrication. When you compare this year to last, last year's number must come from last year's document, not the model's guess about it.

Worked Example: The Supplier Figure, Recalled and Then Grounded

Return to the carbon accountant and the 4,210-tonne supplier figure, and watch the two paths split.

The Recall Path

She types: what were this supplier's 2024 Scope 1 emissions? with no document attached. The model, having never seen the supplier, answers: "Approximately 5,000 tonnes CO2e, which is typical for a manufacturer of this scale." She now has a number that is wrong (the real figure is 4,210), undated in any meaningful sense, sourced to nothing, and dressed as a fact. If she trusts it, her inventory carries a fabricated supplier figure that overstates the footprint by nearly 800 tonnes, and there is no document behind it. When the assurer samples this datapoint and asks for the supplier evidence, the 5,000 has no parent in the file, and the supplier's own response, sitting right there, contradicts it. This is not a small error. It is a fabricated number presented as measured data.

The Grounded Path

She tries again. She pastes the relevant paragraph of the supplier's response into the chat and writes: Below is an extract from our supplier's 2024 emissions response. Answer only from this text. Quote the exact line you rely on, then state the supplier's reported 2024 Scope 1 emissions. If the figure is not stated, reply "not stated in the document."

The model now responds: "The document states: 'Our site Scope 1 emissions for calendar year 2024 were 4,210 tCO2e (third-party verified).' Based on this, the supplier's reported 2024 Scope 1 emissions are 4,210 tCO2e, described as third-party verified." Look at what she has now. The correct figure. A verbatim quote she can confirm against the page in five seconds. The provenance (the supplier's own response) ready for the file. And a bonus the recall path could never give her: the quote surfaced that the figure is third-party verified, a fact about data quality that matters for whether she logs this as primary, supplier-measured data. The grounded path did not just give a better number. It gave a number with a source, a checkable quote, and a clue about its reliability, which together are the difference between a datapoint and a guess.

The Lesson of the Split

Same supplier, same model, same accountant. The only variable was whether the document was in front of the model and whether the model was told to stay inside it. That single choice moved the output from a fabricated 5,000 to a sourced, quoted, verifiable 4,210. Grounding is not a refinement of recall. It is a different mode of operation, and for any number that will reach a disclosure, it is the only acceptable one.

Making Grounding Your Default, Not Your Exception

The mistake professionals make is to treat grounding as something they do for the important numbers and skip for the quick questions. But the quick, ungrounded question is exactly where a recalled figure slips into the workflow unnoticed, because it never felt important enough to check. The discipline is to make grounding the default for anything that touches a disclosure: if a question could produce a figure or a factual claim that might end up in the report, the source document goes in the prompt and the answer comes out of it.

This also reframes what the model is good for. Ungrounded, a model is a confident stranger guessing about your data. Grounded, it is a fast, tireless reader of your documents, pulling the exact line you need and quoting it so you can confirm it. The second is enormously valuable to a discloser drowning in supplier PDFs, utility bills, and prior filings, and it is valuable precisely because it does not ask you to trust the model. It asks you to trust a quote you can see, from a document you provided. That is the whole trick, and it is why grounding turns AI from a disclosure hazard into a disclosure tool.

Hold one principle above all: open-web recall has no place in producing a disclosure number. The model's memory is a fine tool for brainstorming or explaining a concept, and a disqualifying tool for stating what your supplier emitted or what your bill says. For those, there is one rule. Force the model to the document, make it quote the line, and trust your eyes, not its memory.

The Failure Modes Grounding Catches, One by One

It is worth being precise about what grounding protects you from, because each failure mode is a real way an ungrounded answer corrupts a disclosure, and seeing them named makes the technique's value tangible.

The Confident Invention

The most obvious failure is the one in the worked example: the model simply makes up a figure for data it has never seen, delivered in the same fluent voice it uses for things it has. Grounding eliminates this entirely, because a model confined to provided text and required to quote cannot invent a figure that is not in the text without the missing quote giving it away. The invention has nowhere to hide when a verbatim quote is demanded.

The Quiet Paraphrase

Subtler and more dangerous is the paraphrase that drifts. A supplier writes "we estimate our Scope 1 emissions at approximately 4,000 tonnes," and an ungrounded or loosely grounded model reports "the supplier's Scope 1 emissions are 4,000 tonnes," silently promoting an estimate to a measurement and dropping the approximation. That single drift changes how you must label the data in your file, from secondary estimated to primary measured, which is a material difference to an assurer. The verbatim quote catches it instantly, because the word "estimate" survives in the quoted line and forces the honest label.

The Wrong Period

A document often contains figures for several years. Ask ungrounded and the model returns a number with no period attached; ask grounded and quoting, and the period travels with the figure, so you can confirm it matches your reporting year. A right number from the wrong year is still a misstatement, and only the quoted context tells you which year you are holding.

The Blended Answer

The worst failure is the hardest to detect: an answer that is part grounded and part invented, where the model read some of your document and filled the rest from memory. Because you cannot tell which clause is real, the whole answer is suspect. The answer-only instruction is what prevents this, by forcing the model to either stay inside the provided text or decline, never to silently blend. A clean refusal on the missing part is infinitely more useful than a seamless answer that hides the seam.

Across all four, the pattern is the same: the verbatim quote and the answer-only constraint turn invisible failures into visible ones. You are not making the model perfect. You are making its mistakes show up where you can see them, which for a discloser is the entire game, because a visible mistake is one you catch at your desk and a hidden one is the one the assurer catches in the engagement.

Key Takeaways

• Every model answer comes from recall (its training, a compressed average of the open web) or grounding (specific source text you provide). For disclosure, grounding is the only acceptable mode.
• An ungrounded figure is an industry average wearing the costume of a specific fact: it is not your data, it has no date or version, and it is untraceable by design.
• A number you cannot trace is, in an assured disclosure, indistinguishable from a wrong number, because you cannot prove it is right and the assurer only has to find that you cannot.
• The paste-the-source technique has three load-bearing parts: paste the actual text, instruct the model to answer only from that text, and demand it quote the exact line verbatim before answering.
• The verbatim quote is the masterstroke: it is checkable in seconds, it catches paraphrase that drifts from the source, and it means you trust your own eyes rather than the model's summary.
• Ground the documents you live in: the supplier response (preserving measured vs. estimated), the utility bill (the exact kWh), the policy (quote the clause), and the prior report (the real prior figure, not a guessed one).
• In the worked example, the only variable was grounding, and it moved the output from a fabricated 5,000 tonnes to a sourced, quoted, third-party-verified 4,210.
• Make grounding the default for anything that could reach a disclosure, and keep open-web recall out of any reporting number entirely.