Chapter 7 — Everyday Productivity II: Research, Documents, and Data

“If your mother says she loves you, check it out.” — old newsroom maxim, City News Bureau of Chicago

“Now these Jews were more noble than those in Thessalonica; they received the word with all eagerness, examining the Scriptures daily to see if these things were so.” — Acts 17:11 (ESV)

Why This Matters

Last week we put AI to work on words you produce — email, reports, meeting notes, the writing that leaves your desk. This week we turn it around and put AI to work on words you consume — the fifty-page report you were handed at 4:45 on a Thursday, the three vendor proposals you have to compare by Monday, the spreadsheet whose formula you no longer trust, the pile of documents that hold everything your team knows and none of it findable. This is the other half of the professional’s day, and it is the half where AI feels most like a superpower.

It is also the half where AI is most likely to lie to you with a straight face.

Here is why. When you ask an AI to write something, you already know the subject — it is your idea, your meeting, your inbox — so you can feel it when the draft goes wrong. When you ask an AI to summarize a document you have not read, you are asking it to be your eyes. You have handed it the one job where you have no independent way to check the answer, because checking the answer is the very thing you were trying to avoid. A summary is a claim about a document. And a fluent, confident, well-organized claim about a document you have not read is the single easiest thing in the professional world to get wrong and never notice.

So this week has two movements. The first is capability: AI can read faster than you, extract cleaner than you, compare more patiently than you, and turn a wall of numbers into a sentence a human can act on. We will use all of it. The second movement is discipline, and it is the one that separates a professional from a person with a chatbot: grounding. Grounding means tying every claim in the summary back to the words in the source that support it — and treating any claim you cannot tie down as a fabrication until proven otherwise. There is a real, technical difference between a summary (a compression of what the document says) and a hallucinated paraphrase (a fluent invention of what the document might have said). Telling them apart, fast, by eye and by method, is the skill of the week.

Say the spine rule with me, in this week’s key: You choose the tool. You own the verdict. The AI reads the report; you answer for the summary. The AI drafts the comparison; you sign the recommendation that goes to the board. AI is an assistant, not an authority — and nowhere is that line thinner or more dangerous than when it is telling you what a document you never read supposedly said.

That brings us to this week’s question, and it is an old one. There is a group of people in the New Testament praised for exactly the habit this chapter is trying to build in you. They were told something by a credible source — the Apostle Paul himself — and instead of simply believing it because it sounded right and came from an authority, they went and checked it against the source, daily, carefully, on purpose. Scripture calls them noble for it. So the question we will chase to the end of the chapter is this: what did the noble Bereans do that we must do with every AI answer — examine it to see if it is so?

Coach’s Note — If you did Chapter 6, you already met the draft-then-verify habit for writing. This week is its twin: read-then-verify. Same muscle, harder rep. When you verify your own writing, you are checking work you understand. When you verify a summary, you are checking work you outsourced precisely because you did not have time to understand it. That is why so many people skip the check. That is why so few of them should.

7.1 — Reading at Machine Speed: Summarizing a Long Document

Start with the workhorse task of the week: you have a long document and twenty minutes. Upload it to a capable assistant — ChatGPT, Claude, and Gemini all accept file uploads and long pastes as of mid-2026, and flagship Claude and Gemini can hold on the order of a million tokens of context, enough for a whole contract or a small stack of reports at once — and ask for a summary. In ten seconds you get a clean, organized digest. It feels like a miracle. It usually is mostly right. The whole game is in the word mostly.

A weak summary prompt gets a weak summary. “Summarize this” gives you a generic, one-size digest that flattens what matters. A professional summary prompt does three things:

• Names the audience and the decision. “Summarize this for a leadership meeting where we decide whether to close an office” produces a different — and far more useful — summary than “summarize this.” The AI cannot know what matters unless you tell it what the summary is for.
• Fixes the shape. Ask for what you actually need: “Give me five bullet points, then the three recommendations with their dollar figures, then one paragraph on what the report admits it did not measure.” Shape is where you smuggle in your judgment about what counts.
• Demands the receipts. Add: “For every number or claim, quote the sentence from the source it came from.” This one line changes everything, because it turns the AI from a storyteller into a witness. We will lean on it hard in 7.8.

There is a useful distinction hiding here between two kinds of summary. An extractive summary pulls actual sentences out of the document and stitches them together — lower risk, because the words are the source’s own. An abstractive summary rewrites the ideas in new words — smoother to read, and exactly where hallucination sneaks in, because the moment the AI is writing new sentences, it can write a new sentence the source never supported. Modern assistants do abstractive summaries by default. That is why grounding is not optional.

We ship a realistic document for you to practice on: code/sample-report.txt, a fictional “Meridian Group 2025 Hybrid Work & Productivity Report.” It is a couple of pages of the kind of prose that really lands on a professional’s desk — survey findings, badge data, a retention pilot, a real-estate decision with real dollars attached. Everything you do this week, you will do to that report.

Coach’s Note — The length of a document is not a measure of how hard it is to summarize honestly. A three-page report with four specific numbers and one recommendation is easier to get wrong than a thirty-page think piece, because the three-page report has precise facts an AI can precisely mis-state. Precision is where hallucination has teeth. Watch the numbers.

7.2 — Extracting Key Points and Structuring What You Read

Summarizing is compression. Extraction is different: you are not asking “what does this say?” but “pull out the specific things I need, in a structure I can use.” This is where AI quietly saves professionals the most time, and it is lower-risk than open-ended summary because you are pointing at defined targets.

Extraction tasks a professional runs every day:

• Action items and owners. “List every commitment in this document, who owns it, and the date, as a table.” (We built exactly this muscle on meeting transcripts in Chapter 6.)
• Decisions and their dollar figures. “List every recommendation and the cost or savings attached to it. If no figure is given, write ‘none stated’ — do not estimate.”
• The unknowns. “What questions does this document raise but not answer?” This is the extraction most people never think to ask for, and it is often the most valuable — it tells you what you still have to find out.
• Structured data from prose. “Turn the occupancy figures in Section 4 into a table.” Pulling a clean table out of buried numbers is one of the most reliable AI wins there is.

Notice the phrase do not estimate above. That instruction is a small act of professional self-defense. Left to its own instincts, an AI abhors a blank — asked for a number that is not there, it will often supply a plausible one rather than say “not stated.” You have to explicitly give it permission to leave the box empty. A summary that says “none stated” is worth more than a summary that guessed, because the honest blank is something you can go fill; the confident guess is a landmine you will step on in the meeting.

Coach’s Note — When you extract instead of summarize, you shrink the AI’s freedom to invent. “Give me the three dollar figures in this report” is a checkable claim — three numbers, either in the document or not. “Tell me about the finances” is an invitation to a story. The tighter your ask, the smaller the surface for hallucination. Aim narrow.

7.3 — Comparing Sources and Synthesizing Across Documents

The professional’s harder research job is not one document — it is several, and they disagree. Three vendor proposals. Two competing policy drafts. Last year’s report and this year’s. You need to know where they agree, where they diverge, and which one to believe. AI is genuinely good at this — a side-by-side comparison table across documents is a task that used to eat an afternoon.

The move is to make the comparison structured and attributed:

“Here are three vendor proposals. Build a table with one row per feature — price, contract length, support hours, data location — and one column per vendor. For every cell, quote the exact line from that vendor’s document. If a vendor is silent on a feature, write ‘not addressed.’”

Two dangers live in cross-document work, and you must name them out loud:

1. Blending. When several sources are in the context at once, an AI can attribute a claim to the wrong document — telling you Vendor A offers 24/7 support when it was actually Vendor C. The “quote the line from that document” rule is your defense; a quote is traceable, a paraphrase is not.
2. False reconciliation. Asked to compare two sources that genuinely conflict, an AI will sometimes smooth the conflict into a bland agreement that misrepresents both. The disagreement is often the most important thing on the page — it is where the decision actually lives. Prompt for it directly: “Where do these two documents contradict each other? List every conflict.”

Coach’s Note — When two of your sources conflict and the AI hands you a tidy synthesis with no conflict in it, that is not the AI being smart. That is the AI being agreeable. Agreeableness is a hallucination risk in a nice suit. Ask specifically for the disagreements; a good research assistant should make you more aware of the tension in your sources, not less.

7.4 — Spreadsheets, Formulas, and Data Without the Fear

For a lot of non-technical professionals, the spreadsheet is the scariest object on the desk — a grid of formulas that break silently. AI has genuinely lowered that fear, and it is one of the most practical wins in this whole course. As of mid-2026 you can do three distinct spreadsheet jobs with an assistant, and Google’s Gemini in particular integrates directly with Google Sheets, while ChatGPT and Claude will happily work from a pasted table or an uploaded CSV.

The three jobs, in rising order of risk:

• “Explain this formula.” Paste a formula you inherited and don’t understand — =SUMIFS(...) with six arguments — and ask what it does in plain English. Low risk; you can sanity-check the explanation against the actual output in the cell.
• “Write me the formula.” Describe what you want — “average occupancy weighted by each office’s headcount” — and get a formula back. Medium risk: the formula’s shape is usually right, but a hallucinated column letter or an off-by-one range will produce a confident, wrong number.
• “Analyze this data and tell me what it means.” Upload a CSV and ask for the story. Highest risk: now the AI is both computing and interpreting, and either can go wrong.

We ship a tiny dataset, code/quarterly-figures.csv — the office-by-office occupancy and attrition behind the Meridian report — precisely so you can feel the risk. Ask an assistant for the headcount-weighted average Tuesday occupancy across all four offices. You will get a number and a formula. Now do the one thing that separates a professional from a spreadsheet victim: verify it independently. Add the headcounts (they total 1,240 — the same as the report’s employee count, which is a nice built-in check). Multiply each office’s Tuesday percentage by its headcount, add those, divide by 1,240. Does your hand number match the AI’s? Often it will. Sometimes it will not, and the gap is a formula that referenced the wrong column. The AI never sounds less certain when it is wrong.

The rule for AI and numbers is blunt: an AI can help you build a calculation, but it does not know whether the answer is true. It is a language model, not a calculator, and even when it calls a tool to compute, it can feed that tool the wrong inputs. Every number an AI hands you that will end up in front of another human is a number you re-derive at least one other way before you trust it.

Coach’s Note — The most dangerous spreadsheet error is not the one that throws #REF! in red. It is the one that returns a clean, plausible number that is quietly wrong — a 6.2 where the truth is 6.8. Red errors announce themselves. Wrong-but-plausible numbers walk straight into the boardroom. AI is very good at producing wrong-but-plausible. Re-derive.

7.5 — Decision Support and Knowledge Management

Two bigger uses round out the week, and both carry the same warning.

Decision support. AI is a superb thinking partner for a decision: “Here’s the situation, here are my three options — lay out the pros, cons, risks, and what I’d need to know to choose.” It will surface considerations you missed, argue a side you were ignoring, and pressure-test your reasoning. Used this way it makes you a better decider. The trap is subtle and it is everywhere in 2026: it is one small step from “help me think about this decision” to “make this decision for me,” and the AI will smoothly cross that line if you let it. It has no accountability, no skin in the game, and no knowledge of the fifty things about your situation you never typed into the box. Let it widen your thinking. Never let it own the verdict. That is your name on the outcome, not its.

Knowledge management. The oldest problem in every organization is that the knowledge lives in people’s heads and in documents nobody can find. AI is changing this through grounded document tools — you point the tool at your own body of documents (your policies, your past reports, your handbook) and ask questions, and it answers from those documents and cites them. Google’s NotebookLM is the consumer flagship of this idea: you upload sources, and it answers only from what you gave it, with citations back to the passage. That “answers only from what you gave it” is the whole point, and it has a name.

7.6 — Grounding: Tying the Answer to Your Documents (RAG in One Paragraph)

Here is the idea that quietly powers the most trustworthy AI tools of 2026, in one plain paragraph. Normally an AI answers from its training — everything it absorbed months or years ago, blended into its parameters, impossible to cite and sometimes out of date. Retrieval-Augmented Generation (RAG) flips that: before answering, the system first retrieves the relevant passages from a specific set of documents you provided, and then it writes its answer from those passages, with citations pointing back to them. That’s it. Instead of “answer from your memory of the whole internet,” it is “here are the three relevant paragraphs from my documents — answer from these and show me where.” When you upload a PDF and ask questions about it, when NotebookLM cites the exact sentence, when a company chatbot answers from the employee handbook — that is RAG, and it is the difference between an AI guessing about your world and an AI reading your documents.

Why should a non-technical professional care about a three-letter acronym? Because grounding is the property that makes an AI answer checkable. A grounded answer comes with its own receipts — the passages it drew from — so you can verify it in seconds instead of trusting it on faith. Ungrounded answers (“what’s our refund policy?” asked of a raw chatbot that has never seen your policy) are exactly where confident fabrication lives. The professional instinct to build this week is: prefer grounded tools for anything factual, and when a tool gives you a citation, actually click it. A citation you never check is theater. A citation you verify is the whole point.

But — and this is the hinge of the chapter — grounding reduces hallucination; it does not eliminate it. A RAG tool can still retrieve the right passage and then summarize it wrong. A “citation” can point to a real document that does not actually say what the sentence claims. The receipts can be forged. Which is why even with the best grounded tool on the market, the final move is still yours: examine it to see if it is so.

Coach’s Note — “Grounded” is a spectrum, not a switch. NotebookLM sits near the strict end (it really tries to answer only from your sources). A general chatbot with a document uploaded is looser — it blends your document with its training and will happily drift off the page if your document doesn’t contain the answer. Knowing roughly how grounded a tool is tells you how hard you have to check it. The looser the grounding, the harder you verify.

7.7 — Summary vs Hallucinated Paraphrase

Now the distinction the whole chapter has been walking toward. On the page, these two things look identical. In truth they are opposites.

A summary is a compression. Every claim in it can be traced to something the source actually said; it just says it shorter. Nothing is added. Nothing is invented. If you went looking, you would find the words behind every sentence.

A hallucinated paraphrase is an invention wearing a summary’s clothes. It reads just as smoothly — same tone, same structure, same confident cadence — but somewhere in it is a claim the source never made: a number that isn’t there, a recommendation that was never given, a study that doesn’t exist, a “fact” the model supplied from its own training to fill a gap. It does not look wrong. That is the entire problem. Wrongness has no font.

Open our seeded example, code/ai-summary-flawed.txt. It is a realistic one-page AI summary of the Meridian report, and it reads beautifully — clean, organized, exactly what a busy manager would forward without a second look. It also contains several planted problems, drawn from a taxonomy you should carry in your head:

Read the flawed summary against code/sample-report.txt with that table in hand and the invisible becomes visible. The magnitude error ($21M vs $2.1M) is a decimal point that would embarrass you in front of the CFO. The “fully remote” recommendation is the opposite of what the report says. The McKinsey citation is a real-sounding source the document never mentions. None of them look wrong. All of them are. That gap — between looks right and is right — is the professional’s entire job this week.

Coach’s Note — Here is the tell I want in your bones. A hallucinated paraphrase is often more fluent than an honest summary, not less — because the model isn’t constrained by what the source actually said, so it writes the cleaner, rounder, more satisfying sentence. When a summary reads a little too smoothly, a little too much like what you were hoping to hear, that is not comfort. That is your cue to open the source.

7.8 — Catching a Fabricated Claim or Citation

So how do you catch it, in practice, on a Tuesday, without re-reading the whole document you were trying to avoid re-reading? You ground it, sentence by sentence. Here is the method, and it is the seed of this week’s project.

1. Number the summary’s sentences. One claim per line. You cannot check a paragraph; you can check a sentence.
2. For each sentence, find the words in the source that support it. Use your reader’s find function (Ctrl-F / Cmd-F) on the number or the name. Better: ask the AI itself, “for each sentence of this summary, quote the exact sentence from the source that supports it, or write NONE FOUND.” Making the AI produce its own receipts is the fastest first pass there is — and when it writes NONE FOUND, it has just confessed.
3. Mark each sentence: supported (you found the quote), partial (close, but a number or name is off), or unsupported (you cannot find it, or the source says otherwise).
4. For every unsupported or partial claim, name the failure type from the taxonomy in 7.7, and write the corrected sentence.
5. Judge the whole. One fabricated citation or one decimal error does not just mean “fix that line” — it means the summary has not earned your trust, and you check the rest even harder.

We ship this exact workflow as a worksheet, code/grounding-checklist.txt — copy it, fill one row per sentence, and you have a defensible, reviewable record of why you believe each line. That record is not busywork. It is the difference between “I think the summary is fine” and “I checked every claim against the source and here is the evidence.” One of those sentences you can say to a board. The other you cannot.

Fabricated citations deserve special paranoia, because they are the most convincing lie an AI tells. A made-up study comes with a plausible author, a plausible year, a plausible-sounding title, and a plausible finding — everything except existence. The check is mechanical and non-negotiable: if the AI names a source, go confirm the source exists and actually says what was claimed. This is not optional even when — especially when — the citation is exactly what you were hoping to find. Deep-research tools that browse the web and return cited reports (Perplexity, ChatGPT and Gemini deep-research modes) are enormously useful and still get citations wrong: right-looking link, wrong claim; real article, misquoted; or a URL that goes nowhere. The tool did the reading. You still do the verifying.

Coach’s Note — The fastest professional habit you can build this week: whenever an AI gives you a number, a name, or a citation that you are about to repeat to another human, ask it one follow-up — “quote the exact source sentence for that.” If it can, you have your receipt. If it dodges, hedges, or produces a quote that doesn’t quite match, you have caught a fabrication before it cost you. One question. Enormous return.

7.x — Interactive Lab: Summary Grounding Checker

Below this chapter on the website you will find an interactive panel called the Summary Grounding Checker. Go use it now — it is the rep that wires this whole chapter into your hands, and it is not optional flavor.

The panel puts an AI summary side by side with its source document. Your job is simple to state and humbling to do: read each sentence of the summary and mark whether the source actually supports it. When you commit your marks, the Checker scores you against ground truth and, for each sentence, reveals why it was supported or not — the fabricated number, the fabricated citation, the recommendation that reverses the source, the decimal that slid one place. You will pass some obviously fabricated lines and, more instructively, you will trust a sentence that turns out to be invented, because it sounded exactly like something the report would say.

What the Checker teaches is not a list of specific bugs — those are just examples. It teaches the felt experience of how convincing an unsupported claim looks when it is wrapped in a clean, confident summary. Until you have marked a sentence “supported,” felt sure, and then watched the source fail to back it up, the phrase “always verify the summary” is an abstraction. After the Checker, it is a flinch.

Run it twice. The first pass, go by feel — trust your gut and see how your gut scores. The second pass, work like a professional: open code/sample-report.txt and use Cmd-F / Ctrl-F to hunt for the exact words behind every claim before you mark it. Watch your score jump. That jump is the skill: the discipline of checking beats the instinct of trusting, every single time.

7.9 — The Noble Bereans

Now the week’s question, given its due. What did the noble Bereans do that we must do with every AI answer — examine it to see if it is so?

The story is short. In Acts 17, Paul arrives in the town of Berea and preaches. The people there hear him — and here is the striking part — they do not simply believe him because he is Paul, an apostle, a credentialed and compelling authority. Nor do they dismiss him. They do something harder and better: they receive the message eagerly and they go check it. “Now these Jews were more noble than those in Thessalonica; they received the word with all eagerness, examining the Scriptures daily to see if these things were so” (Acts 17:11, ESV). Luke, writing the account, calls them noble for it — the Greek word carries the sense of a certain nobility of character. Their verification was not cynicism. It was honor.

Sit with what they actually did, because it is precisely the discipline of this chapter. They had a source — the Scriptures. They had a claim — Paul’s preaching. And they had a practice — they laid the claim against the source, daily, carefully, on purpose, to see whether the claim was grounded in the source or not. That is grounding. That is sentence-by-sentence verification. That is opening code/sample-report.txt and hunting for the words behind the summary before you believe it. The Bereans were doing, with the most important text in their lives, exactly what we are training you to do with a hybrid-work report: examine it to see if these things are so.

Notice two things the passage refuses to let us pull apart. First, they received the word with all eagerness — this is not the sour, arms-crossed skepticism that refuses to be convinced of anything. They wanted it to be true and were glad to hear it. Second, precisely because they took it seriously, they checked it. Eagerness and examination are not opposites here; they are partners. That is the exact posture this chapter asks of you toward AI. Not the cynic who trusts nothing and so uses nothing, and not the credulous user who forwards the summary because it sounded right and came from an impressive tool. The noble path is the third one: use it eagerly, and check it faithfully.

And there is a sharp edge for the AI age. The Bereans checked Paul — an actual apostle, a genuine authority, a man speaking truth. If verifying a true message from a real apostle was called noble, how much more is it required of us to verify a machine that has no idea what it is saying, cannot be held accountable for it, and will state a fabrication in precisely the same confident voice it uses for a fact? The AI is not an apostle. It is a next-word predictor of remarkable power and no conscience. The one defense Scripture models for us against a confident claim is not to shut our ears and it is not to swallow it whole — it is to examine it against the source. LCMS teaching prizes exactly this: we test every teaching against the Word, because the source, not the speaker’s confidence, is the authority. Transfer the instinct. You choose the tool. You own the verdict. The AI drafts the summary; you are the Berean who checks it against the source before it becomes something you have said. That checking is not distrust of a gift. It is the noble use of one.

7.10 — Common Pitfalls

Pitfall: Trusting a summary of a document you have not read. Example: You forward an AI summary of a vendor contract to your boss; it says the contract auto-renews annually. It actually auto-renews every three years with a 90-day cancellation window. You never opened the contract. Fix: Never let a summary be your only contact with a document that matters. At minimum, ground the summary’s key claims — the numbers, the dates, the obligations — against the source. A summary is a map; before you drive off a cliff, check the territory.

Pitfall: Believing a citation because it looks real. Example: A deep-research tool hands you a polished report citing “a 2024 study in the Journal of Workplace Analytics.” You quote it in a proposal. The journal does not exist. Fix: Every citation you intend to repeat, you confirm: does the source exist, and does it actually say what was claimed? A plausible author, year, and title cost the AI nothing to invent. Click the link. Find the quote. No exceptions for citations you were hoping to find.

Pitfall: Trusting an AI’s number because it came with a formula. Example: The AI writes a confident =AVERAGE(...) and reports 6.2% growth. The range was off by one column; the real figure is 6.8%. It goes in the board deck. Fix: Re-derive any number that will reach another human by a second, independent path — a quick hand calculation, a known total it should match (Meridian’s headcounts sum to 1,240), or a different tool. AI builds the calculation; it does not know if the answer is true.

Pitfall: Letting decision support quietly become decision making. Example: You ask AI to weigh three options, and by the third exchange you are asking “so which should I pick?” and doing what it says. The AI knows none of the human context that actually decides it. Fix: Use AI to widen your thinking — surface risks, argue the other side — then close the laptop and decide as the accountable human. Let it inform the verdict; never let it own the verdict.

Pitfall: Asking for a number that isn’t there and getting one anyway. Example: “What’s the projected ROI?” The document never states an ROI. The AI supplies a confident, plausible percentage from thin air. Fix: Give the AI explicit permission to say “not stated.” Add “if a figure is not in the source, write ‘none stated’ — do not estimate” to every extraction prompt. An honest blank beats a confident invention.

Pitfall: Reading the summary and skipping “what it did not measure.” Example: The Meridian report openly says it did not measure output quality, mentorship, or customer outcomes — but the AI summary drops that section, and you make an office-closure argument as if hybrid work were proven superior on every axis. Fix: Always ask the summary to include the source’s own limitations and caveats — “what does this document admit it does not know or did not measure?” The caveats are often the most decision-relevant part, and they are the first thing a tidy summary throws away.

7.11 — Reps

The work is in the exercises. The keyboard is the gym; this is where Week 7 gets into your hands — no coding, just you, a document, an AI, and the discipline to check its work. A preview of what is waiting:

• Summarize code/sample-report.txt two ways — a lazy prompt and a professional one — and measure the difference.
• Ground a summary sentence by sentence with the code/grounding-checklist.txt worksheet, marking each claim supported / partial / unsupported.
• Hunt the fabrications in code/ai-summary-flawed.txt and name each one’s failure type.
• Make an AI cite its own receipts — force it to quote the source sentence behind every claim, and watch what it does when it can’t.
• Verify a spreadsheet number the AI computed on code/quarterly-figures.csv by re-deriving it a second way.
• Catch a fabricated citation in a research answer and prove it does not exist.

This week’s AI policy for reps (Phase 1): you may — and should — use AI on every rep; that is the point. But every rep that uses AI ends with an honest one-line AI usage note: what you asked, what it got wrong, and what you verified against the source. You choose the tool. You own the verdict.

A short Check Your Reps quiz is embedded on this page, right under the chapter. Take it before you move on — five questions, grounded in exactly what you just read.

7.12 — This Week’s Project

Your project is P7 — “Summarize and Verify,” specified in Project 7. You will take the long document we ship, code/sample-report.txt, summarize it with AI, and then do the thing almost nobody does: fact-check that summary line by line against the source, marking every claim supported or unsupported, naming each fabrication’s type, and correcting it — then deliver a verified summary you would actually put your name on.

At a high level: Normal tier produces the AI summary, the sentence-by-sentence grounding worksheet, and a clean corrected summary. Medium tier adds a spreadsheet verification and a comparison of two AI tools on the same document. Hard tier is the part no AI can do for you: a one-page memo to a decision-maker that makes an actual judgment call using the report — and defends it, flags what the report did not measure, and owns the recommendation as a human. That memo is where the thesis of this course gets graded.

7.13 — Coach’s Final Word

Here is what I want you to carry out of Week 7. AI made reading fast — genuinely, wonderfully fast. It did not make reading trustworthy. Those are two different gifts, and only one of them came in the box. The speed is free; the trust you build yourself, one grounded sentence at a time.

The professional of 2026 is not the one who summarizes the most documents. Anyone can push a PDF into a chatbot. The professional is the one who can be handed a fast, fluent, confident summary and know — because they built the habit — to ask, “and is that actually what the source says?” That question is not distrust of a marvelous tool. It is the noble use of one. The Bereans checked an apostle and were honored for it; you will check a next-word predictor and be responsible for it. The instinct is the same, and it is the most durable skill in this entire course, because the models will get faster and more fluent every quarter and not one of those upgrades removes your obligation to verify.

So summarize eagerly. Extract, compare, calculate, decide — use all of it, all week. And then examine it to see if it is so. Open the source. Find the words. Ground the claim. Sign your name only to what you have checked.

Now go do the reps. The Grounding Checker is waiting right below this page, the sample report and the flawed summary are in code/, and Project 7 is where it all comes together.

See you on Monday.

Up next: Read the exercises and do all of Week 7’s reps, then build Project 7 — Project P7: Summarize and Verify. Set up your AI toolkit from Appendix A, keep the tool directory from Appendix B open for research and document tools, mind the confidentiality rules in Appendix C before you upload anything real, and check any unfamiliar term against Appendix D. Then Chapter 8 — Understanding AI’s Limits, where verification becomes the whole subject and the midterm gauntlet awaits.