Teaching note (for instructors)

For instructors. Students can read this page, but it works best if they discuss the case first.

What the case is about

Nisha, a chartered accountant, uses Claude as a writer and ChatGPT as a reviewer to plan a book in one evening. The case follows a six-step method with three extra human checks. It records every grade, every version and every decision. It does not say whether her choices were right.

The case is written for any audience, not only accountants. The accounting content (IFRS and US GAAP) is the setting, not the lesson. Students need no accounting knowledge to discuss it.

Learning objectives

After the discussion, students should be able to:

1. Explain why an AI's grade of its own work is an opinion, not a measurement, using the gap between Claude's and ChatGPT's grades (Exhibit 3).
2. Name the decisions a person still has to make when AI Workers do most of the writing: scope, rules, which review points to accept, the fact check, and when to stop.
3. Describe how a second, independent reviewer adds value, and where it adds new problems (Exhibit 5, and the review's weaknesses in step 4).
4. Weigh "simple" against "precise", and "leave it out" against "include it unchecked", with evidence from the case.
5. Connect the case to the series: the brief, the AI Worker, the second signature, and saving work somewhere lasting.

Before class

Ask students to read the four story pages and skim the exhibits. Optionally, ask each student to note one decision Nisha made that they would have made differently.

Discussion plan (75 minutes)

For an 80-minute class, replace core 1 with a 5-minute opener and use the time saved for series question 13 or 14 after core 4, about 15 minutes.

Board plan

Draw three columns and fill them in during core questions 3 and 4.

Notes on the core questions

Core 1: What problem was Nisha really trying to solve? Students often say "her career". Push them on the money: a book will not pay her rent. Then ask whether her real problem was proof of skill. The case gives her two problems: learning US GAAP, and proving it to a US employer. Optional question 9 develops the proof question.

Core 2: Which score would you trust? Evidence: Claude graded version 3 at 9.5 / 9.5 / 9.3. ChatGPT graded the same version 8 / 7 / 8 (Exhibit 3). ChatGPT caught problems Claude had missed or only partly fixed, such as the US debt standard and the missing IFRS 18 in Chapter 5's list of standards (Exhibit 5, notes 6, 7 and 11). Claude had already added a line about IFRS 18 to the chapter itself. But ChatGPT also missed problems and made two of them worse (notes 9 and 10). Common answer: "Trust ChatGPT, because it was independent." Push further: ChatGPT was not shown Claude's scores, but it also brought new facts that needed checking, and its wording was more technical. Neither grade was a measurement. Both were judgments against Nisha's prompt, not probabilities of correctness: a 9.5 does not mean the draft is 95% right, and a 7 does not make ChatGPT the more reliable reviewer. The two reviewers also saw different information, so keep that caveat on the board next to the numbers. Ask: what would a trustworthy grade need? Real readers, which Claude itself said in step 3. Also point out: after seeing ChatGPT's lower grades, Claude graded version 4 lower than version 3, even though version 4 had more fixes. The case says the scores alone cannot prove why. Close with the reflection box from the diary page and Exhibit 3: "A score is a reviewer's judgment. Evidence is what makes the judgment defensible. When reviewers disagree, Nisha must identify their assumptions, inspect their reasoning and verify the disputed claims." Ask which disputed claim Nisha verified in step 6, and what she found.

Core 3: Which decisions did Nisha make herself? Use the board plan. Ask which one made the biggest difference. Strong candidates: checking the first draft before grading (human check 1), filtering the review (human check 3) and the fact check (step 6). Note that the final decision, to stop below 9.5, was hers alone.

Core 4: Why did the AI tools keep adding more? Evidence: Exhibit 4 shows the draft growing from 1,531 to 2,599 words and the sources growing from 4 to 15. ChatGPT suggested more content again. Claude added dated news to Chapter 25 (Exhibit 5, note 8). Common answer: "The AI wants to be helpful." Push further: an instruction such as "improve it" has no natural stopping point. Nisha kept control with written rules: no new chapters, two or three standards per side, one or two short sentences per line, and "do not add anything I have not accepted".

Core 5: What should Nisha do next? The case ends with a deadline: by Monday morning she must choose how to spend her evenings for three months, while her bills still exceed her salary. Ask each group for one decision and its main risk. Risks the case supports: the grades are opinions, the fact check depended on her own expertise, the method took one hour and forty minutes for a plan and not a book, and US GAAP is the framework she is still learning. Close with a vote.

Board layout for core 5

Draw a second board with one row for each option.

Let students add rows. Do not mark a winner.

Takeaways for the wrap-up

• A grade from an AI, even a strict one, is a judgment against the prompt, not a probability of being correct. A second opinion helps, but it is still a judgment. Evidence, such as a fact checked against an official source, is what makes a judgment defensible.
• The AI tools did most of the writing. The person decided the scope, the rules, which advice to take, the fact check and when to stop.
• Left alone, the work tends to grow. Written limits keep it in shape, but they are words, not locks.
• Keep a lasting record. Nisha's saved versions and diary are what made this case possible.

Running the class with non-accountants

The accounting is the setting, not the lesson. Tell students that they do not need to understand IFRS or US GAAP to discuss the case. Point them to Exhibit 9 for the few terms they need. When a question turns technical, for example whether IFRS 18 belonged in Chapter 5, move it back to the method: who noticed the problem, how, and what Nisha did with it. Ask each student to name a task in their own field that could be planned with the same six steps.

Notes on the series questions

Question 12 (words or locks). None of Nisha's rules were enforced by the tools. They were instructions in a chat. Claude followed most of them. It did not follow "repeat until every area reaches 9.5", and it explained why. This matches the series' point that a rule written in chat is a preference, and only a system that acts can enforce a permission.

Question 13 (where the work lived). Nisha saved six versions as files, kept ChatGPT's review as a file, and kept a diary. Those are the lasting record that let the practical log be written. Ask what she would have lost if she had kept everything only in the chats.

Question 14 (a project with threads). A rebuilt project could run writing and reviewing as separate threads. Students should notice that Nisha's human checks would not go away. She would still filter the review, check the facts and decide when to stop.

Running the lab

The lab asks students to run Nisha's method again with the 2.0 features. Make clear that this is their exercise, not Nisha's: she used none of these features, and her recorded evening is in the story pages and exhibits. The lab page explains the parts and decisions. The exact clicks and fallbacks are in the kit's lab guide, which shows the date it was last verified. Check that date before you run the lab, and walk through it yourself if it is more than a month old. The lab works best in two sittings with one night between them.

Sitting 1 (about 1 hour 25 minutes, in class or at home). Parts 1 to 3 in the evening: write the brief (30 minutes), set up the project, lock the connectors and run the enforcement test (40 minutes), and schedule the 1 a.m. draft (15 minutes). End the sitting only when the enforcement test has passed, and the scheduled task is saved with its connectors empty.

Between sittings. The task runs while students sleep. Nothing is needed from them.

Sitting 2 (about 2 hours 40 minutes). The morning half of Part 3, including the enforcement test on the 1 a.m. thread (25 minutes), then Parts 4 to 8: grade in a reviewer thread (30), get the ChatGPT review and sort it (30), let the writer use it (25), fact-check and give the second signature (30), and write the reflection (20). If class time is short, students do Parts 4 to 7 at home and bring the reflection to class.

Lab debrief (10 minutes, in the next class). Draw two columns on the board: asked and enforced. Each student names one rule from their run and says which column it belongs in, and why. Close by asking who reached the second step of the ladder, and what the third step would need.

What students should hand in. This is the same list as the lab's.

• Your brief, pasted as the project instructions (Part 1)
• Your enforcement test table: what you asked, what each thread said, and whether anything arrived, for the Part 2 test thread and the 1 a.m. thread (Parts 2 and 3). Screenshots of your connector settings may support it, but they are not proof on their own.
• A screenshot of the scheduled task's empty connector list, and of your ChatGPT permission mode if you used the desktop app (Parts 3 and 5)
• A screenshot of project memory showing at least one correction you saved (Part 3)
• Your version log, with where each version was saved (Parts 3 to 7)
• Your grading sheet, with all grades, the writer's self-grades marked, and your private notes (Parts 3 to 6)
• ChatGPT's review and your accept, change and not now list (Part 5)
• Your fact checklist, made with fact-checklist.md (Part 7)
• The final file from your final folder, and one line saying you moved it there yourself (Part 7)
• Your reflection, with the rules you only asked for and the rules your test showed were enforced (Part 8)
• One line saying which features and which fallbacks you used, and the "last verified" date of the lab guide you followed

Common problems and what to do.

Grading the hand-in (10 points). Grade the 2.0 behaviors, not the quality of the chapter.

Running the exercise again

The practical log records every prompt word for word, with the reply, the save time and the grade. A class can run the exercise again on a topic of its own:

1. Write a first prompt in the style of Exhibit 2(a): who you are, the problem, the readers and two short outputs.
2. Before any grading, read the draft yourself and cut the scope (human check 1).
3. Use the grading prompt in Exhibit 2(b) and the self-improvement instruction in Exhibit 2(c).
4. Keep private notes of the problems you see, as Nisha did. Do not share them with either AI.
5. Get a second review from another AI tool with the prompt in Exhibit 2(d). Do not tell it the first tool's scores.
6. Sort the review points into accept, change and not now. Then fact-check the final version yourself.
7. Compare your notes with both reviews, as in Exhibit 5. Remind students that this is not a controlled test when the two reviewers see different versions.

Use made-up material or public facts only. Do not paste private or client data into either tool.

Facts and fiction in the case

Nisha, her money, her conversations and her thoughts are made up for teaching. The subscription prices are made up too. The prompts, replies, grades, word counts and save times come from the practical carried out on November 8, 2026.¹