Level 200: Using gen AI well · Module 200.3

Two coworkers look down at an open tabbed binder in a café booth.
Open to the page

Grounding it in your documents · Plain gear

Reviewed and published, version 1.0.0. As of 18 Nov 2026. Not a credential.

You will be able to:

Big Widget Company, and every Big Widget document in this lesson, is made up.

Giving the AI the right pages

Think of a friend helping with your homework. If they answer from memory, they might be right or wrong. If you hand them the right page of the textbook first, they have a much better chance.

That is grounding. You ask the AI to answer from documents you give it. There are two ways:

The scientists who named it in 2020 said it was still hard for an AI to show where its answers came from.1 In one small test, they gave their AI newer documents, and its answers changed to match. They did not have to rebuild the AI.1 That was a science test on Wikipedia, not on a company's papers.

So grounding helps in two ways. The AI can use your rules, not a guess. And when your rules change, the AI can use the new ones.

Handing it the page is not enough

Here is the wrong idea: "If I attach the document, the answer comes from the document." Not always.

The AI may skip parts of the page. In 2023, scientists gave AIs a question and a pile of 10, 20 or 30 documents. Only one held the answer. They moved that one around.2 The AIs did best when the answer was near the start or the end of the pile, and worse when it was in the middle.2 Sometimes, with a big pile, one AI did worse than with no documents at all.2 Those were the AIs of 2023. Newer ones were not tested. But the lesson lasts: having the page is not the same as reading it.

The AI may add things. NIST, a US government science agency, has a name for an AI saying false things in a confident voice: confabulation.3 NIST says this includes answers that wander away from what the AI was given.3 So the AI can quote the right rule and then add a bit that is not in it.

The page itself may be wrong. If the page is old, or two pages disagree, the AI can read them perfectly and still be wrong. More on that below.

At Big Widget Company: Lee and the returns rule. Lee works in dealer support. Lee gives the AI the returns rule. It says: "This policy covers all returns, including returns from dealers. Unused parts may be returned within 30 days of delivery with the original invoice. Custom-built widgets cannot be returned." A dealer wants to return an unused part after 40 days. Lee asks three times and gets three answers:

All three sound like the rule. Only one is. Here is how to tell.

Ask for the exact words

In the course’s request step, ask for an exact quote and its location, or “not found” if the documents do not answer. Then open the source, read around the quote, and check that it supports the claim.
The retrieval paper identifies provenance as an open problem (s1). The course’s procedure asks for evidence and checks each citation and claim against the source, following NIST’s verification guidance (s4).
Figure data as text

Quote, point, verify

  1. Quote and point (course step) — ask for exact words, the document version or date, and the page or section. Ask for “not found” if there’s no answer.
  2. Open the source — find the exact quote and read around it. Check that it supports the claim. [s4]

Objective: 200.3-b. Text equivalent of figures/quote-point-verify.svg.

Ask the AI to show its work. Every time, ask it to:

  1. Quote the exact words from the document, in quotation marks.
  2. Say where they are: which document, which date, which section.
  3. Say "not found" if the document does not answer the question.

For example: "Answer only from the returns rule. For each thing you say, quote the sentence and give the section. If the rule does not say, write 'not in the rule'."

Asking does not make the AI right. It makes the AI easier to check.

Check the words yourself

Here is another wrong idea: "If the answer gives a source, the source must say it." Not so. NIST warns that an AI can make up reasons or sources that seem to back up its answer, and that these can fool people into trusting it too much.3 A source in an answer is just another thing to check. NIST says to check sources and citations in AI answers.3

Four steps:

  1. Open the real document. Not the AI's copy. The real, current one.
  2. Find the quote. Is it really there, word for word?
  3. Read around it. Does the next sentence add a "but" or an "except"?
  4. Match each claim to the quote. Does the quote really say what the answer claims?

Step 4 catches the sneaky ones. The AI might quote the 30-day sentence perfectly, then add the 15% fee. The quote is real. The fee is not in it. Cross the fee out, and do not pass it on.

Check every part of a long answer, not just the start. Remember: AIs can miss the middle.2

When the documents are the problem

The last wrong idea: "When the answer is wrong, it is always the AI's fault." Sometimes the documents are the problem:

NIST says good information can be traced back to where it came from, and makes clear when it might go out of date.3 So every document an AI reads should have a named owner and a date to check it again.

A true story: Air Canada's chatbot

In November 2022, someone in a customer's family died. The customer asked the chatbot on Air Canada's website about cheaper fares for people travelling after a death in the family.4 The chatbot said they could ask for the lower price after the trip, within 90 days of the day the ticket was issued.4 But the chatbot's answer linked to Air Canada's own page about these fares. That page said you could not ask after the trip.4 One website, two different answers.

The customer took Air Canada to a tribunal in British Columbia, Canada.4 Air Canada said it should not have to answer for its chatbot, as if the chatbot were on its own.4 In February 2024, the tribunal said no. The chatbot is part of Air Canada's website, and Air Canada is responsible for everything on its website.4 The tribunal said the customer did not have to double-check the chatbot against other pages.4 Air Canada had to pay.4 (The case is called Moffatt v. Air Canada, after the customer. We leave the name out of the story because the lesson does not need it.)

What it teaches: the company owns both answers, the page and the chatbot. So the company must make sure the things it publishes agree. That is not the customer's job.

What it does not tell us: the tribunal did not say how the chatbot came up with its answer.4 So we do not know if the documents or the AI went wrong. The honest answer is "unknown", and we do not guess. It is also just one small case, in one place.4

Who fixes it, and when to pass it up

When an answer is wrong, ask one question first: do the documents look right?

An owner is a person with a name. If you do not know who it is, write "owner unknown" and ask your manager. Do not guess.

Working out exactly why the AI got it wrong is for Level 300. Your job is to spot it and send it to the right person.

At Big Widget Company: Lee, two weeks later. In this made-up company, Ana owns the returns rules. Tom runs the AI.

The three yeses

The Foundation names three moves: Get to Yes, Stay at Yes and Recover to Yes.5 A Foundation gate gives one of four answers: "go ahead · only this far · wait for a person · no".5 Here is how this course uses them for answers from your documents.

Keep the evidence, not the secrets. Lee's note to Tom has the question, the answer, the quote and the real page. That note stays in Big Widget's own work files. A Foundation receipt is different. It "holds no private content".5 It never carries "prompts, outputs, payloads, personal data".6 So the dealer's details and the AI's answer do not go in a receipt.

My checklist

Words to know: grounding, retrieval, confabulation, quote, source, owner, pass it up (escalate), gate, receipt, Get to Yes, Stay at Yes, Recover to Yes.

Want to learn more? (optional) The University of Washington's free lessons Modern-Day Oracles or Bullshit Machines?, by Carl T. Bergstrom and Jevin D. West, explain why an answer that sounds sure is not proof: <https://thebullshitmachines.com/>.

Sources

  1. s1 Patrick Lewis et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Paper. Checked 2026-11-18 by source-checker role, separate Claude session 8bfb2ae6 (not the writer), first check at 1bb203d; re-confirmed by separate session a1770ea6 at 3c40af2 (round 2). ↩ ↩
  2. s2 Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni and Percy Liang (2023). Lost in the Middle: How Language Models Use Long Contexts. Paper. Checked 2026-11-18 by source-checker role, separate Claude session 8bfb2ae6 (not the writer), first check at 1bb203d; re-confirmed by separate session a1770ea6 at 3c40af2 (round 2). ↩ ↩ ↩ ↩
  3. s4 Chloe Autio et al. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1). Standard. Checked 2026-11-18 by source-checker role, separate Claude session a1770ea6 (not the writer), round-2 re-check at 3c40af2. ↩ ↩ ↩ ↩ ↩ ↩
  4. s3 Christopher C. Rivers, British Columbia Civil Resolution Tribunal (2024). Moffatt v. Air Canada, 2024 BCCRT 149. Record. Checked 2026-11-18 by source-checker role, Devin session f9ae423f (did not write, revise or red-team this module), remaining-sources check, 18 Nov 2026. ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩
  5. s5 AiGovOps Foundation (n.d.). AiGovOps Foundation Manifesto. Official doc. Checked 2026-11-18 by source-checker role, separate Claude session a1770ea6 (not the writer), round-2 re-check at 3c40af2. ↩ ↩ ↩
  6. s6 AiGovOps Foundation (n.d.). AiGovOps Foundation Practice (home page). Official doc. Checked 2026-11-18 by source-checker role, separate Claude session a1770ea6 (not the writer), round-2 re-check at 3c40af2. ↩

Next: the quiz