Try each exercise first, then open its model answer.
Checking the work · Workbook
Three exercises, one for each thing this module teaches. Everything here is made up, set at Big Widget Company. Do not use real names, customers or documents from your own work. Write short answers. "I don't know yet" is an honest answer if you also say who you would ask.
Exercise 1 (100.8-a)
Four checks on one AI answer. Luis in planning asks the company's approved AI tool for a short note for his director. The director will use it to decide whether to move a production line to a second shift. The AI's note says:
"Overtime costs rose 18% last quarter, from $2.1 million to $2.6 million (Finance report Q3, page 4). Industry guidance recommends a second shift once overtime passes 15% (Smith & Lee, Plant Staffing Handbook, 2022). Recommendation: move Line 3 to two shifts."
- For each of the four checks (source, number, sense, second person), write one thing Luis should do.
- Who should the second person be, and what should they open?
- Luis is busy. He says: "It's from the approved tool, and checking will take longer than writing it myself." What do you tell him?
- Until the checks are done, what happens to the note?
Model answer
- The four checks:
- Source. Open the Finance report for Q3 and go to page 4. Is the overtime figure there? Then look for the Plant Staffing Handbook. Does it exist? Does it really recommend a second shift at 15%? If Luis cannot find it, the claim is not checked and comes out of the note.
- Number. Work it out again. From $2.1 million to $2.6 million is a rise of $0.5 million, which is about 24%, not 18%. Either the numbers or the percentage are wrong. Find which in the report.
- Sense. Does this fit what Luis knows? Was there a one-off rush order last quarter? Is Line 3 the line with the overtime, or is it another line?
- Second person. This drives a big decision, so a named person checks it too, and they open the report themselves, not just Luis's note.
- A good answer names someone who did not write the note and knows the figures, such as a finance analyst who owns the Q3 report. They open the report and the handbook, not just the summary. Asking the AI "are you sure?" is not a second check, and the AI never signs off.
- A good answer points to Johnson v. Dunn (2025). The firm had a rule to check AI results, but the lawyers did not open the cases. The judge said even a spot check would have quickly found the fake citations. Skipping it led to public reprimands, and the firm paid another law firm to check more than 2,400 citations. Here, recomputing one percentage takes a minute and has already found an error. An approved tool can still be wrong.
- It waits. Luis does not send it as fact. If something cannot be checked in time, he marks it "not checked" and the director does not rely on it for the decision.
Exercise 2 (100.8-b)
Three questions before you repeat it. Read these three claims. They are made up.
- A. A sales slide from an AI vendor: "Our assistant answers customer questions correctly 92% of the time."
- B. A colleague in a meeting: "Everyone knows AI writes better emails than people now."
- C. A news story about an AI chatbot "beating humans" on a test, from a study several years old.
For each claim:
- Ask "out of how many?", "measured how?" and "over how long?". Write one specific question for each.
- Say what you would write in a meeting note if nobody can answer them.
Last, in one or two sentences: does "92%" in claim A prove that someone measured it well?
Model answer
A. "Correctly 92% of the time."
- Out of how many? 92% of how many questions? Which kinds of question: easy ones, or the hard ones our customers ask?
- Measured how? Who decided what counted as "correct"? Was "I don't know" counted as correct? (In TruthfulQA, a non-answer counted as truthful, and usefulness was scored separately.) Who ran the test: the vendor, or someone independent?
- Over how long? When was it tested, and on which version? NIST says lab results may differ from real-world use, and AI should be tested regularly while in use.
B. "Everyone knows AI writes better emails."
- Out of how many? Which emails, and whose? Compared with which people?
- Measured how? Better by what measure: fewer mistakes, faster replies, happier customers? Who judged?
- Over how long? When was this true, and with which tool?
- A good answer notices there is no evidence at all here, only confidence. In Johnson v. Dunn, the judge said a senior lawyer's big statement "came with no citation or other basis". Ask what it is based on.
C. "Beats humans on a test."
- Out of how many? How many questions, and how many humans? In TruthfulQA, the AI's 58% was out of 817 questions, but the 94% "human" figure was one person, on 250 of them, with the internet. Different denominators.
- Measured how? Were the AI and the people tested on the same questions under the same conditions? Were the questions picked on purpose to be tricky?
- Over how long? The study is several years old. Which models did it test, and is the result still true today?
The meeting note. For each one: "Claim not checked", and the name of the person who will find the evidence. Do not repeat it as fact.
Does "92%" prove anything? No. A percentage makes a claim sound measured. It does not show who measured it, how, out of how many, or when.
Exercise 3 (100.8-c)
Real customer, or not? Big Widget Company's marketing team wants testimonials for a new product page. Sort each idea into fine or fake, and give a one-line reason.
- Ask AI to write three short quotes from "happy customers", with first names and job titles.
- A real customer, Ana, sent a long, rambling email praising the delivery team. AI shortens it. Ana agrees to share it and approves the final wording.
- Use AI to write a paragraph in the company's own voice: "We build widgets to last."
- A real customer, Raj, said the product "works fine". AI rewrites it as "Raj says it doubled his output", and nobody asks Raj.
- A colleague says: "AI tools like Rytr got let off by the FTC, so AI reviews must be legal now." Is the colleague right?
Last, in one or two sentences: why do AI-written "customer" quotes fool people?
Model answer
- Fake. No real customers said these words. Readers are told real customers had real experiences, when no one did. That breaks the FTC rule on two counts: the reviewer does not exist, and the experience did not happen. The FTC said AI-generated reviews are covered.
- Fine (this course's reading of the rule, not wording from it). Ana is real, she had the experience, she agreed, and the meaning stays hers. A person on the team still signs off the page.
- Fine. It is marketing in the company's own voice. It does not pretend to be a customer.
- Fake. Raj is real, but the quote misstates what his experience was. Under the FTC rule, any one of the three is enough: that the person exists, that they had experience with the product or business, or what that experience was.
- No. The FTC set aside the Rytr order because its complaint had not claimed that Rytr knew, or should have known, that customers used the reviews to break the law, or even that they were published. The case was about the toolmaker. The same order says people who use AI to break the law should be held accountable. A business that writes or posts fake customer reviews is what the review rule describes. And rules change, so check what is in force today before relying on it.
Why it fools people: people trust a testimonial because they believe a real customer is speaking from real experience. An AI-invented quote borrows that trust while no one had that experience. Using AI is not the problem by itself. Pretending a customer said it is.