Try each exercise first, then open its model answer.
How it writes · Workbook
Two exercises, one for each thing this module teaches. Everything here is set at Big Widget Company, a made-up company. The people, tools, documents and answers are made up too. Use only what the lesson taught. You do not need to look anything up.
Exercise 1 (100.3-a)
At lunch, your coworker Sam says: "The AI tool is amazing. It must think through the whole answer first, like I would, and then type it out. So by the time it shows me something, it has already checked it."
Sam's tool is built on an autoregressive model, the kind of model this lesson is about (GPT-3 is one example). Write three to five short sentences back to Sam. Your reply should:
- Say how a model like this makes its answer, using the word token.
- Say what the model reads each time it picks the next token, and that there is a limit to how much it can read.
- Say why this way of writing does not, by itself, check the facts.
Model answer
A good reply covers all three points, in plain words. For example:
"I can't tell you what goes on inside it while it works; the lesson doesn't settle that. But the writing itself happens one small piece, a token, at a time. Each time, it reads your question plus what it has written so far, works out which token is likely to come next, adds one, and goes again. There's a limit to how much text it can read at once: for GPT-3 it was 2,048 tokens. And it picks tokens because they are likely, not because they are checked. Likely isn't the same as true, so the answer still needs checking."
What to look for:
- One token at a time. The reply says the model adds one small piece at a time, each one chosen as a likely next piece. A reply that says the model writes the whole answer in one go, or that writing it this way means it has already been checked, has kept the misconception.
- No penalty for honest limits. Do not mark a reply down for saying the lesson does not settle what happens inside the model, or that some tools add separate steps around the model (such as looking up documents first). Full marks keep those steps apart from the token-by-token writing, and do not treat asking the same model again as an independent check.
- What it reads. The reply says the model reads the request plus what it has already written, and that each new token becomes part of what the next one is based on. Full marks also mention that the amount it can read at once has a limit.
- Likely, not checked. The reply says that picking a likely token is not a fact check, so a model working normally can still write something false.
- Not about every AI. The reply talks about this kind of model. It does not claim that every AI system writes this way, or say how a particular app shows the words on screen.
Exercise 2 (100.3-b)
Mei in purchasing asks an AI tool: "Why did the cost of steel brackets go up last quarter? List your sources." The tool answers:
Steel bracket costs rose 18% last quarter for two reasons. First, our main supplier raised its prices in July (Source A: Supplier price notice, July). Second, freight costs doubled because of port delays (Source B: Freight Costs Review, Q3, page 12). Costs are expected to fall next quarter (Source B, page 14).
Mei opens each source. Here is what she finds:
- Source A is a real notice in the purchasing files. It says the supplier raised bracket prices by 6% from 1 July.
- Source B cannot be found. Nobody in finance or logistics has heard of a Freight Costs Review, and there is no such file.
- The 18% figure does not appear in Source A or anywhere else Mei can find.
Mei's manager wants the explanation for a report to the board tomorrow.
Answer these questions:
- The answer reads smoothly and lists sources. Using what NIST says, explain why it can still contain things that are false, even if the AI tool is working normally.
- Go through the answer claim by claim: the 18% rise, the supplier price rise, the freight costs doubling, and the costs falling next quarter. For each one, say what Mei should do with it.
- Mei's coworker suggests: "Just ask the AI if it's sure about Source B." Is that a good check? Say why or why not, and say what Mei could do instead.
Model answer
1. Why a smooth answer can still be false. The tool writes by picking likely next pieces of text, and likely is not the same as true. NIST calls confidently stated false content confabulation (often called hallucination) and says it is a natural result of how these models are designed, not a sign that the tool is broken. NIST also warns that AI output can include made-up citations that make an answer look trustworthy. A source list is just more text the model wrote, so it may point to nothing real.
2. Claim by claim.
| Claim | What Mei found | What she should do |
|---|---|---|
| Costs rose 18% | No source supports it | Take it out, or ask finance (for example, Priya) for the real figure and use that, with its source. |
| Supplier raised prices in July | Source A is real and says 6% from 1 July | Keep it, rewritten to match Source A exactly: a 6% rise from 1 July. |
| Freight costs doubled | Source B does not exist | Look for a real source. If none supports it, take the claim out. Do not just delete the fake source and keep the claim. |
| Costs will fall next quarter | Source B does not exist | Same as above: take it out unless a real source supports it. |
3. Asking the same AI. No. Asking the same AI "are you sure?" is not an independent check: it can answer just as confidently whether or not Source B exists. Mei has already done the real check by looking for the source. Because this is going to the board, she should also ask a named person, such as Priya in finance, to check the numbers. If something still cannot be checked by tomorrow, she should write "not checked" next to it and tell her manager, rather than guess. Her manager can then decide: go ahead with only the checked parts, or wait.
What to look for:
- Not a sign of a fault. Answer 1 says made-up content can come from a tool working normally, as a result of how the model writes, not only from a broken tool. An answer that says "the tool must be faulty" has kept the misconception. Do not mark an answer up or down for saying how often this happens: the lesson's sources give no single rate for every model and task. An answer that states a rate for this tool without saying what was measured, on which model and when, has gone beyond the evidence.
- The claim, not just the citation. Answer 2 removes or replaces unsupported claims. An answer that only deletes Source B, or swaps in some other real document on a nearby topic, has missed the point.
- Match the source exactly. Answer 2 changes the supplier claim to what Source A actually says (6% from 1 July), rather than keeping the answer's wording.
- A check that is independent. Answer 3 says the same AI cannot check itself, and names opening the source and asking a real person as the checks.
- Honest gaps. Full marks say what to do when a claim cannot be checked in time: say "not checked" and take it to someone, rather than guessing.