# Lesson 3: Working with AI: Prompting, Context and Verification

*Companion-podcast transcript • Sarah & Kiffer*  
*Office Hours episode to listen to after working through the lesson*

---

**Sarah:** In one consulting experiment, about eighty-four percent of consultants without AI reached the right answer on a particular task.

**Kiffer:** Compared with sixty to seventy percent of those with AI. And on the tasks inside the tool's frontier, the consultants with AI worked about a quarter faster and produced work rated more than forty percent higher in quality.

**Sarah:** So the same tool helped them on some tasks and left them worse off on another. How were they supposed to know which was which?

**Kiffer:** The researchers chose that task so it would lie beyond what the tool could do. Your own tasks don't arrive labelled, so you have to find that edge yourself.

**Sarah:** Welcome back to Office Hours. I'm Sarah.

**Kiffer:** And I'm Kiffer. This is Lesson three, Working with AI: Prompting, Context and Verification. Sarah has the first question.

**Sarah:** Question one. What puzzled me when I started was that the same assistant could explain a tricky statistical idea really well and then miscount the words in a short paragraph. Why would the easy task be the one it fails?

**Kiffer:** Because easy and hard describe how a task feels to a person. For a person, counting words takes patience, and explaining a confidence interval takes years of training. A language model works with tokens, which are words or pieces of words, and it writes by predicting likely continuations. Explaining draws on patterns it has seen many times. Counting needs exact tracking of every item, which prediction handles poorly unless the tool runs code. So the frontier follows how the model works, and our own sense of difficulty is a poor guide to where it lies.

**Sarah:** And the consulting experiment showed why that matters. Inside the frontier, consultants with AI finished about twelve percent more tasks, worked about a quarter faster, and produced work rated more than forty percent higher in quality. On the one task outside it, about eighty-four percent of consultants without AI reached the right answer, compared with sixty to seventy percent of those with AI.

**Kiffer:** The result I find most useful is who gained the most inside the frontier. The consultants who'd scored lowest on a baseline task improved by forty-three percent, compared with seventeen percent for the higher scorers. So on those tasks, the tool narrowed the gap between weaker and stronger performers.

**Sarah:** That worries me a little. If the people who gain the most are the ones with the least experience, they might also be the least able to tell when they've crossed outside the frontier.

**Kiffer:** It's a reasonable worry, and I'd treat it as an open question, since the results we've quoted for the outside task describe the consultants as a group. It's one reason a frontier map records a check for every task, including the ones inside. There's one more limit worth knowing. The researchers chose the outside task so that it would lie beyond what the tool could do. Your own tasks don't arrive labelled, so you have to find the frontier yourself with known-answer tests.

**Sarah:** And here's a problem with known-answer tests that I think students run into. The obvious source of a known answer is a worked example in a textbook. But a well-known textbook problem may have been in the model's training text.

**Kiffer:** That's a good catch. If the model may have seen the very problem, a right answer tells you less than it seems to. So change the problem. Use new numbers, a different setting or a different outcome, and work out the answer yourself first. It's the same idea as validating a screening tool. You test it against a reference standard, on cases like the ones it will meet in real use.

**Sarah:** Okay, so here's what I'd have done. I change the numbers in a rate problem, run it three times in new conversations, and the tool gets it right all three times. I'd put that task inside the frontier and stop recomputing.

**Kiffer:** Let's test that plan. Suppose the tool actually gets this kind of problem wrong one time in five, at random. What's the chance it passes all three of your runs anyway? Take a few seconds.

*(Pause)*

**Sarah:** If it's right four times out of five, the chance of three right in a row is 0.8 times 0.8 times 0.8, which is 0.512. So about half the time, a tool that's wrong one time in five would pass my test. Okay, three runs aren't enough to justify dropping the check.

**Kiffer:** Right. Three runs will catch a tool that fails often, or one that makes the same mistake every time. They can't show that a tool is rarely wrong. Even ten clean runs would leave about a one in ten chance of missing an error rate of one in five, since 0.8 to the power of ten is about 0.11.

**Sarah:** So what do the three runs give me?

**Kiffer:** A rough sorting, which tells you where to spend your checking time. A task that fails even once in three runs is clearly at the edge or outside. A task that passes all three is probably inside, and it keeps its check on the map. For any number that someone else will rely on, that check is recomputing it.

**Sarah:** Question two. Can a better prompt stop a tool from making things up? A lot of students add a line like "Do not invent sources" and assume that's taken care of.

**Kiffer:** It helps less than people hope. In the lesson's fictional Cedar Valley case, the co-op student Jordan wrote a slide prompt that told the assistant not to invent statistics, and that same prompt produced a fabricated reference. A prompt changes what the model has to work with. Unless the tool can search or retrieve documents, the model has no step where it checks its claims against the world, and no wording in a prompt adds one. The part of a prompt that does the most for accuracy is the material you supply, like the program facts, because the model then takes those facts from your text.

**Sarah:** What about asking it to name a source for every claim?

**Kiffer:** I'd keep that in any prompt that produces factual claims, because it makes errors faster to find. A named source is something you can look up, and a request for a page number or a quoted sentence takes you straight to the passage. Just remember that the named source can itself be invented, as it was on Jordan's slide.

**Sarah:** And if I'm suspicious of a reference, can I ask the assistant whether it's real?

**Kiffer:** That's reading vertically. You're asking the source to vouch for itself. The Avianca case is a good illustration. According to the judge's sanctions opinion, one of the lawyers asked ChatGPT whether one of the invented cases was real, and it said yes, and that the case could be found in the major legal databases. The better habit comes from Wineburg and McGrew's work with professional fact checkers. Leave the chat and look the reference up somewhere independent, like PubMed or a library database.

**Sarah:** There's a subtler problem I want to raise. The prompt itself can carry an error. Sharma and colleagues found that assistants often repeated mistakes contained in a user's question.

**Kiffer:** Here's a made-up prompt that shows it. "Write three bullets for clinic managers explaining why social prescribing reduces loneliness among older adults." What's the problem with that prompt? Take a few seconds.

*(Pause)*

**Sarah:** It assumes the answer. It tells the tool that social prescribing reduces loneliness and asks only for reasons. So the tool will write reasons, and maybe a convincing statistic, for a claim nobody has checked.

**Kiffer:** And this claim needs care. When Bickerdike and colleagues reviewed the evidence in 2017, they found fifteen evaluations of social prescribing, all at high risk of bias, and concluded that the evidence wasn't enough to judge whether it works or whether it's good value for money. A tool that's handed the conclusion will tend to write it up with confidence.

**Sarah:** So how would you rewrite it?

**Kiffer:** Ask the open question, and say what you need. Something like, "What does published research say about whether social prescribing changes loneliness in older adults? List findings from controlled evaluations separately from claims in program descriptions, say where each can be checked, and tell me if the evidence is weak or mixed." Everything that comes back still goes to its source, because the tool is recalling the literature, and recall sits outside the frontier.

**Sarah:** Or better still, paste in the abstract of a recent review and ask it to work only from that.

**Kiffer:** Yes. That turns a recall task into a summarizing task, which moves it inside the frontier. Then the check is reading the summary against the abstract.

**Sarah:** Question three. If a number in an output matches the number in the source, does the claim check out? Let me try this one, because I suspect I'd have got it wrong.

**Kiffer:** Here's a made-up output. An assistant drafting a newsletter paragraph writes, "A major review of a hundred and forty-eight studies found that people with strong social relationships live fifty percent longer." The review is real. It's the 2010 meta-analysis by Holt-Lunstad and colleagues. You open the article. What do you conclude?

**Sarah:** The abstract says it pooled a hundred and forty-eight studies, with about three hundred and nine thousand participants, and found a fifty percent increased likelihood of survival for people with stronger social relationships. A hundred and forty-eight studies, fifty percent. Both numbers match, so I'd keep the sentence.

**Kiffer:** Before you keep it, compare the wording as well. What would "live fifty percent longer" actually mean?

**Sarah:** That people live half again as long. So if a typical lifespan were eighty years, people with strong relationships would live to a hundred and twenty. Okay, that's clearly wrong. Nobody is claiming that.

**Kiffer:** So what does the source say?

**Sarah:** It reports an odds ratio of 1.5 for survival. The odds of still being alive at the end of follow-up were about one and a half times as high for people with stronger relationships, and follow-up averaged seven and a half years. That's about surviving over a period. The paper doesn't give any figure for years of life gained.

**Kiffer:** Right. So the figure is real, the source is real, and the claim has changed. That's misattribution, and it's why a primary-source check compares four things: the wording, the figures, the population and the date. The wording matters in a second way. A reader could hear "live longer" as a promise that relationships add years to life. The studies measured people's relationships and followed them over time, so the pooled result is an association, and the authors themselves note that cause is hard to establish.

**Sarah:** So matching numbers are only the start of the check. Can we try one where we do the arithmetic ourselves?

**Kiffer:** Yes. Here's a harder one, where the numbers and the words disagree in two places. The first-wave screening counts are from the lesson, and the rest is made up. Of the six hundred and twenty older adults who screened positive in the first wave, four hundred and thirty-four agreed to meet a community connector. Jordan asks the assistant for one sentence for an internal memo. It writes, "Seventy percent of older adults screened in the first wave agreed to meet a connector, about two for every thousand older adults in Cedar Valley." Look at the first half of that sentence. Is it right? Take a few seconds.

*(Pause)*

**Sarah:** First the number. Two-thirds of six hundred and twenty is about four hundred and thirteen, so four hundred and thirty-four is a bit more than two-thirds. Worked out exactly, four hundred and thirty-four divided by six hundred and twenty is 0.7. So seventy percent is the right number.

**Kiffer:** Seventy percent of whom?

**Sarah:** Ah. The sentence says of older adults screened. Two thousand four hundred and eighty were screened. Seventy percent is a share of the six hundred and twenty who screened positive. Of everyone screened, it's four hundred and thirty-four divided by two thousand four hundred and eighty, which is 0.175, or seventeen and a half percent.

**Kiffer:** So recomputing the number alone would have passed a wrong sentence. Now the second half, two for every thousand older adults.

**Sarah:** Cedar Valley has about forty-six thousand residents aged sixty-five and older. Four hundred and thirty-four divided by forty-six thousand is about nine per thousand. Two per thousand is what you get if you divide by all two hundred and ten thousand residents. A quick check agrees. Older adults are a bit more than one in five residents, so dividing by everyone makes the rate about four and a half times too small.

**Kiffer:** Good. And there's a third layer. Even nine per thousand divides by every older adult in the region, while the first wave runs in only twelve of the twenty-four clinics. People whose clinics aren't offering the program yet couldn't have been referred. I'd leave the regional rate out and write the sentence with its setting.

**Sarah:** So, in the twelve first-wave clinics, four hundred and thirty-four of the six hundred and twenty older adults who screened positive, or seventy percent, agreed to meet a connector.

**Kiffer:** That's the sentence. A prompt habit helps here too. When you ask a tool to report a figure, ask it to name the numerator and the denominator in words. That puts the question, a share of whom, in front of you every time.

**Sarah:** Question four. Is aggregate data always safe to paste into an approved tool? The Cedar Valley pilot allows aggregate data in its enterprise assistant, so it seems simple.

**Kiffer:** Usually it is. The trouble comes with small numbers. Here's a made-up table Jordan might want help laying out. It shows positive screens by clinic, age group and sex. Most cells hold dozens of people, but one cell, for a clinic in a small town, shows two men aged eighty-five and older. Can this table go into the pilot's assistant as it stands? Take a few seconds.

*(Pause)*

**Sarah:** The table looks aggregate, but that one cell points to two particular men in a small town, and someone who knows the town might work out who they are. So that cell is personal information, and a document takes the most protected class it contains. The table can't go into the assistant as it stands.

**Kiffer:** That's how I'd read it. And when you're unsure, you treat it as the more protected class until someone with authority decides. Your organization may also have its own rules for small counts, and those rules come first.

**Sarah:** So how does Jordan still get help with the layout?

**Kiffer:** The layout doesn't need the real numbers. Jordan can give the tool the column headings and a few made-up numbers of the same form, get the table looking right, and put the real figures in afterwards, outside the tool. That keeps the benefit of the tool and keeps the data where it belongs.

**Sarah:** The other belief I hear a lot is, I pay for my account, so my chats are private.

**Kiffer:** Paying doesn't settle it. As of 2026, OpenAI says it may use content from its paid consumer plans, like Plus and Pro, for training unless the user turns a setting off, and Anthropic asks its consumer users to choose. The terms change with an institutional account, under a contract the organization has reviewed, like the version of Microsoft Copilot Chat that SFU students sign into with their university account. Even then, approval for general study and work doesn't extend to personal health information.

**Sarah:** The federal government draws a similar line. The Government of Canada's guide on generative AI says public servants must not put personal information into publicly available online generative AI tools.

**Kiffer:** The same guide asks public servants to take responsibility for the content these tools generate, and to check it against trusted sources. So the federal guidance covers both habits we've been talking about, deciding what goes in and checking what comes out.

**Sarah:** Okay, now the part where we disagree. How much checking is enough when you're studying?

**Kiffer:** The verification routine from the lesson puts private study at the lower level of stakes. You read critically, compare explanations with your course materials, and check any fact before you rely on it. I think that's the right level. Explaining standard concepts is usually inside the frontier, and if you checked every study explanation at a primary source, you'd lose most of what makes these tools useful for learning.

**Sarah:** My worry is that studying is when you're least able to spot an error. When I'm learning something new, I don't know enough yet to notice when an explanation is slightly off. A wrong figure on a slide might get caught by a colleague. A wrong idea I pick up while studying goes into my head and comes out later in work other people read.

**Kiffer:** That's the strongest case for checking more, and I take it seriously. Where I'd push back is on what the check needs. Comparing with your course materials doesn't require expertise, because you're checking against a reference that someone qualified has already checked. A tutoring prompt that asks you questions and waits for your answers adds another way to find gaps.

**Sarah:** But my course materials don't cover everything I ask about. The moment I ask a follow-up the textbook doesn't address, I've lost my reference standard.

**Kiffer:** True. For those answers, I'd treat what the tool says as a lead, something to look up or ask your instructor about before you use it.

**Sarah:** So where do you land?

**Kiffer:** I'd keep private study at the lower level, with the course materials as the reference standard and follow-up answers treated as leads. Raising every explanation to the moderate level would cost a lot of time, and for standard concepts I think the gain would be small.

**Sarah:** I'd still check more than Kiffer would, especially for material I'm meeting for the first time. But we agree on the practical rule. When a claim you learned from an AI tool moves out of your notes and into anything someone else will read, it gets the checks for that new level, starting with the primary source.

**Kiffer:** Let's pull it together with three things to take away.

**Sarah:** First, a known-answer test with a few runs is good at finding tools that fail often. Use changed versions of problems you know, and keep a check on the map for every task, including the ones that pass.

**Kiffer:** Second, a prompt supplies what the tool can't know, and no prompt removes the need to verify. Give it the facts, ask for locations you can check, and keep your own assumptions out of the question.

**Sarah:** And third, verifying a claim means comparing its meaning as well as its numbers. For every figure, ask what it's a share of, and which population and period it describes.

**Kiffer:** If you'd like more practice, rework the referral memo example with your own made-up numbers, and say the numerator and denominator in words for every figure before you accept a sentence.

**Sarah:** Next time, it's Lesson four, AI for Learning and Study, where we look at the evidence that AI can both strengthen and weaken learning, and how to set up tutoring conversations that keep you thinking.

**Kiffer:** Take care, everyone.

**Sarah:** See you in Lesson four.