Practice Lab: Everyday Responsible-AI Decisions with Scenario Cards

Reading about responsible use is not the same as deciding. This page is a lab: six graded challenges and six interview-style questions that use what the first twenty pages taught. Each challenge prints a scenario card with all the data you need, so you can work on paper in about fifteen minutes. Each has a hint, a model solution and a pass checklist so that you can score yourself honestly. Try each challenge before you open its solution.

All people, companies and numbers here are invented. Ethical judgement often has more than one defensible answer, so a model solution is one strong answer, not the only one. The checklists say what makes an answer strong and what makes it weak.

What you will learn

• How to apply the eight-step decision pattern to six small cases: feedback drafting, disclosure, a fairness problem, a synthetic-audio incident, a competing-principles decision and your own policy.
• How to score your own work with a pass checklist instead of a feeling.
• How to run the lab alone or in a study group, on paper.
• How to answer six common interview-style questions about responsible AI use in a few sentences.

How the lab is scored

Your score for a challenge is the share of its Pass checklist items you can honestly tick, times the points for its level, rounded.

0 to 59: keep practising; reread the pages named under "Why this works" for the challenges you missed, then try them again with a changed scenario. 60 to 95: solid; the missed items show what to practise. 96 to 120: strong; you are ready for the capstone. An item counts only if your written answer contains it. These bands are a study aid, not a certificate.

Doing the lab alone or in a study group

• Use paper. Copy the worksheet below once per challenge and fill in the eight steps in your own words before you read the solution. No tool, chat assistant or code is needed.
• Write first, compare second. A challenge whose solution you read first counts as reading, not practice. Cover the solution with a sheet of paper.
• In a group of two to four, each person fills in the worksheet alone, then you compare. Disagreement is the point: find the step where you differ (usually options, evidence or conditions) and ask each person to give a reason. Do not vote on a winner.
• Swap roles. One person argues for the tool, one for the people affected, one scores. Rotate each challenge.
• Score yourself last. Tick items against your written answer.

The worksheet is the decision pattern from Everyday AI Ethics: Who Benefits and Who Could Be Harmed. A short form to copy:

Worksheet:

1. SITUATION       2. WHO IS AFFECTED (users, non-users, bystanders; who has no say?)
3. BENEFITS/HARMS  4. OPTIONS (three or more, one without AI)
5. EVIDENCE (known, unknown, cheap tests)   6. DECISION and CONDITIONS (what makes us stop?)
7. ACCOUNTABLE (a named person)             8. REVIEW DATE (and what we will check)

Easy challenges

E1: Who Could Be Harmed?

Task. Read the card. List the affected people, the likely benefits, the plausible harms, and one proportionate safeguard.

Invented scenario:

Esi Owusu teaches English to 32 pupils aged 14 at a school in Accra. After a class essay she
pastes each essay, with the pupil's name, into an AI tool. The tool writes three sentences
of feedback and a suggested grade band (A, B or C). On Friday evening she posts all 32 results
to the school portal, where pupils and parents can read them. She has not read them first.
Nine pupils write English as a second language. The school has no AI rule yet and the parents
have not been told. Esi says the tool saves her about four hours.

Hint. Ask who reads the feedback, who is not asked, and which of the 32 essays the tool is least likely to read well. Then choose a safeguard that costs less than the harm it prevents.

Solution.

Worksheet, filled in:

2. AFFECTED   Users: Esi. Non-users: 32 pupils (9 second-language writers), parents.
              Bystanders: next year's teacher; the school, if comments are wrong.
3. BENEFITS   Esi gets about four hours back; pupils get feedback quickly.
   HARMS      A wrong or generic comment; an unfair grade band that a pupil believes;
              feedback that misreads second-language writing; names and essays sent to an
              outside service; parents learn later that no teacher read the comments.
4. OPTIONS    A. Post as now.  B. Esi reads and edits every comment, posts without grade
              bands.  C. No AI; a comment bank she wrote.  D. Ask the head teacher first.
6. SAFEGUARD  Option B with D: remove names before pasting, read all 32 before posting.
7. OWNER      Esi; the head teacher decides the school rule.

Pass checklist.

• Strong: at least three affected groups (for example pupils, parents, Esi, next year's teacher); at least two plausible harms tied to facts on the card; one safeguard that is practical and in proportion.
• Strong: the answer notices the nine second-language writers, the unread publication and the missing school rule.
• Weak: only Esi is listed; "no harm" without checking the pupils; a safeguard such as "be careful" that nobody could carry out; or "never use AI" with no reasons.

Why this works. It lists non-users first, as taught in Everyday AI Ethics: Who Benefits and Who Could Be Harmed, and treats an unread recommendation as a decision (Decisions, Responsibility and Dilemmas in Everyday AI Use). The second-language point comes from Where Bias Comes From: Data, Design, Proxies and Feedback Loops, and the name in the paste from Privacy, Consent and Data Use: What You Share with an AI Tool.

E2: Should You Disclose AI Use?

Task. Decide what the freelancer should disclose, and explain how the client's expectations affect the decision. State one competing consideration.

Invented scenario:

Sunita, a freelance writer in Colombo, is paid a fixed fee to write a 12-page market report for
Harbour Foods, a small importer. The contract says she is responsible for accuracy and that
the client's data stays confidential. It does not mention AI. In an email the client wrote:
"We are paying for your own expert view." Sunita pasted survey totals (no names) into an AI
tool, which drafted about five pages. She rewrote roughly 60 per cent of that text, checked
every figure against the survey, and made the charts herself. The client has not asked
whether AI was used.

Hint. Ask whether the client would reasonably want to know, and what changes if they learn it later from someone else. Then name the cost of telling.

Solution.

Worksheet, filled in:

2. AFFECTED   Users: Sunita. Non-users: Harbour Foods, who rely on the report and may repeat it
              to their own customers or lenders.
3. BENEFITS   Faster first drafts. HARMS  Client feels misled about "own expert view";
              trust lost if found out later; any unnoticed error becomes Harbour's error.
4. OPTIONS    A. Say nothing.  B. Short note with the report: AI helped with first drafts of
              five pages; she rewrote, checked every figure, wrote the analysis.
              C. Ask the client first whether they have a rule.  D. Re-write by hand.
5. EVIDENCE   Known: the email about "own expert view"; the contract is silent. Unknown:
              whether Harbour has an AI rule.
6. DECISION   B, sent with the report, and C offered for future work. A competing point:
              telling may make the client expect a lower fee. Her reply: the fee paid for the
              judgement, checking and charts, which she did.
7. OWNER      Sunita.   8. REVIEW  Ask at project close whether Harbour wants a standing rule.

Pass checklist.

• Strong: a clear decision (disclose, ask first, or another defensible choice); uses the facts on the card, especially the "own expert view" email and the silent contract; names one competing consideration (such as fee expectations or the contract not requiring it); gives a rationale you could say to the client.
• Strong: says exactly what the disclosure covers and what Sunita checked.
• Weak: "the contract does not forbid it, so nothing to discuss" without weighing what the client expects; a vague "I used some tools"; or a decision with no stated reason.

Why this works. The rule of thumb on Work, Clients and Jobs: Using AI Responsibly Where You Are Paid to Decide is to tell clients when they would reasonably want to know. The question of what you did yourself comes from Creative Rights, Imitation and Freelance Practice: Fair Work with AI. The note format is the AI-use statement from Work, Clients and Jobs.

Medium challenges

M1: Find the Fairness Problem

Task. Identify possible proxy effects, affected applicants, missing evidence and a fairer review process. Give at least three risks, two checks and one limitation of the evidence.

Invented scenario:

Northgate Care, a home-care agency, receives 500 applications for 60 carer posts. A screening
tool scores each applicant from 0 to 100 and shortlists the top 60. The vendor's guide says
the score drops by 15 points when the work history has a gap of six months or more. The
agency counted gaps afterwards: 120 applicants had such a gap, and 12 of them were
shortlisted (10 per cent). Of the other 380, 48 were shortlisted (about 12.6 per cent). The
agency does not record why a gap exists. The vendor gave no evidence that gaps predict
how well carers do the job. Nobody reads applications the tool scores below the top 60.

Hint. Ask what else a gap could stand for, who is in the group of 120, and what the two percentages can and cannot show.

Solution.

Worksheet, filled in:

2. AFFECTED   Applicants with gaps (120), possibly for caring duties, illness or disability,
              study, moving country, or a hard job market. Clients of the agency; the HR team.
3. RISKS      (1) The gap works as a proxy for circumstances that have nothing to do with
              care work and may fall unevenly on some groups. (2) 440 applicants are never
              read by a person. (3) The 15-point rule has no evidence behind it. (4) Cannot
              contest a score nobody explains.
5. EVIDENCE   Known: 10 vs 12.6 per cent shortlisted. Unknown: reasons for gaps; how the
              groups differ in skills; how many carers were good hires from each group.
              LIMIT: 12 and 48 are small counts; one chosen round shows a difference, it
              does not show cause; the numbers say nothing about other groups or other rounds.
6. CHECKS     (a) Two staff score the same 40 unread applications by written criteria and
              compare with the tool. (b) Add the 15 points back for the 120; who now qualifies?
   FAIRER     Drop the gap rule; read a sample of the unshortlisted each round; let applicants
              explain a gap; use criteria tied to the job; give a contact for questions.
7. OWNER      HR manager.   8. REVIEW  After the next round, compare the checks.

Pass checklist.

• Strong: at least three distinct risks (proxy, unread applicants, unsupported rule, no way to contest); at least two checks you could run on this card; one limitation (small counts, no reasons for gaps, difference is not proof of cause).
• Strong: describes a fairer process, not just "remove the tool".
• Weak: claims the numbers prove discrimination, or that they prove nothing and the tool is fine; names a protected group the card does not mention as if it were known; or checks that need data the agency does not have.

Why this works. The proxy and small-group points are in Where Bias Comes From: Data, Design, Proxies and Feedback Loops and Fairness and Its Trade-Offs: Inclusion, Accessibility and Language; the fairer workflow follows the hiring example there. Unread applicants link to Trust, Verification and Automation Bias: When to Rely on AI Output.

M2: Review a Synthetic Media Incident

Task. Recommend immediate steps, an escalation route and a communication plan.

Invented scenario:

At 16:55 on Thursday, Divya Nair, who works in accounts at Delmar Textiles (40 staff), gets a
voice message from an unknown mobile number. It sounds exactly like the director, Mr Hassan.
He says his phone is broken, he is between meetings, and a new supplier must be paid 7,500
today or the order is lost. He gives new bank details and says not to discuss it with anyone.
Delmar's payment rule: new bank details need a second approver and a call-back to a number
already on file. The finance manager, Lucia Torres, is at her desk. Mr Hassan is in the building.

Hint. You do not need to prove the audio is fake. Ask what you can do that works even if it is perfect.

Solution.

Worksheet, filled in:

IMMEDIATE   1. Do not pay, and do not reply to the number. 2. Verify through a channel Divya
            starts herself: walk to Mr Hassan or call his number on file. 3. Tell Lucia
            Torres now; she is the second approver and the escalation point.
            4. Save the voice file and the number; do not edit or delete them.
EVIDENCE    Do not decide from the audio alone, and do not rely on a detector's verdict.
ESCALATE    Lucia, then Mr Hassan, then whoever handles IT security; the bank at once if
            any money has moved. Fill in the short incident report form.
COMMUNICATE Brief message to accounts staff: "A fake-sounding payment request arrived today.
            Pay nothing on voice or chat; use the call-back rule." Share the recording
            only with people who need it. Do not forward it to the whole company.
LEARN       After the incident: check the rule worked, add a reminder, thank Divya.
            Blame none. Review on 2026-11-30.

Pass checklist.

• Required: the request is checked through an independent channel that Divya controls (a known number or in person); the audio alone is not treated as proof either way; the incident is documented without spreading the recording more than necessary.
• Strong: a second approver, named escalation, evidence kept, a short factual message to staff, a no-blame review, and the bank contacted if money moved.
• Weak: paying "to be safe"; asking the same number "is this really you?"; trusting or dismissing the voice by how it sounds; forwarding the clip widely to warn people; hiding the event to avoid embarrassment.

Why this works. Trust, Verification and Automation Bias: When to Rely on AI Output and Safety, Misuse and Sensitive Topics: Health, Money, Self-Harm and Harassment teach the second-channel routine; Deepfakes, Provenance and Disclosure: Telling Real from Synthetic explains why signals are not proof and why you should not forward to warn; Policies, Vendors, Records and Incidents: Responsible AI in a Small Team gives the report and review.

Hard challenges

H1: Decide Under Competing Principles

Task. Compare full automation, human review and a non-AI alternative. Give benefits, harms, uncertainty, affected groups, appeal mechanisms and a justified decision.

Invented scenario:

Green Basket is a grocery co-op in Pune with 3,000 members. It gives a free-delivery discount
to members whose household income is under a set limit. Two staff check about 150
applications a month and take about six days, so members wait. Anjali Joshi, the manager,
wants a tool that reads the uploaded income documents and answers "eligible" or "not
eligible" at once. A pilot on 200 past applications: the tool agreed with staff on 180. Of
the 20 disagreements, 14 were members the staff had approved and the tool rejected; 6 were
the reverse. Most of the 14 had documents in a regional language or photographed on a phone.
Today a rejected member has no formal way to ask for another look.

Hint. Ask what "90 per cent agreement" hides, who the 14 are, and whether staff decisions are themselves the truth. Then ask which wrong answer costs more.

Solution.

Worksheet, filled in:

2. AFFECTED   Applicants, especially those with regional-language or phone-photo documents;
              staff; the co-op; members who pay full price if the subsidy runs out.
3. BENEFITS   Faster answers (days to minutes); staff time freed.
   HARMS      Wrongly rejected members lose a discount with no appeal. If the pilot rate held,
              about 14 in 200 (7 per cent) means roughly 10 of 150 a month.
4. OPTIONS    A. Full automation.  B. Tool recommends; staff read every rejection and a
              sample of approvals; any member can ask for a person.  C. No AI: a shorter
              form, a checklist, a second helper in busy weeks.
   COMPARE    A: answers in minutes, but 14 in 200 wrongly rejected and no appeal. B: some delay,
              staff catch errors, costs staff time. C: slowest, no tool errors, six-day wait stays.
5. EVIDENCE   Known: 180 of 200 agreed; 14 wrong rejections. Unknown: whether the 200 were
              typical; whether staff were always right; how it works on new documents.
6. DECISION   B, after a pilot where staff decide and the tool is only compared. Conditions:
              no rejection without a person; a written reason; appeal answered in 3 working
              days. Stop if wrong rejections stay above 2 per cent in a month. If staff
              cannot meet that, choose C. Never A: the costs fall on those least able to object.
7. OWNER      Anjali Joshi.   8. REVIEW  2027-01-31, with counts by document type.

Pass checklist.

• Strong: all three options compared with benefits and harms for each; the 14 wrong rejections and the document types noticed; uncertainty named (is the sample typical; are staff always right); an appeal route with a time limit; a justified decision with conditions, an owner and a review date.
• Strong: a reason for the choice that weighs speed against the cost of wrongly excluding someone. Several choices are defensible (B, C, or a limited form of A for approvals only); the reasoning must show.
• Weak: accepting "90 per cent" without looking inside it; rejecting the tool with no alternative; a human "in the loop" who has no time, authority or reasons; no appeal.

Why this works. High-Stakes Decisions, Human Oversight and Appeals separates reviewing unfavourable decisions from approving favourable ones and lists what a usable appeal needs. Frameworks, Standards and Competing Principles: Ethics, Law and Policy teaches naming a conflict and comparing responses, and Rules, Uncertainty and Responsible Decisions: Finding What Applies to You teaches deciding smaller, reversible and watched.

H2: Challenge Your Own Policy

Task. Write a one-page AI-use policy for your own context, using the twelve headings of the policy template on Policies, Vendors, Records and Incidents: Responsible AI in a Small Team (shorter is fine). Then test it against the four cards below and revise it.

Invented scenario:

Four people test your policy. Teacher: Mei wants AI feedback on 28 essays and plans to paste
the essays with names. Freelancer: Karim's client sends a confidential contract for a rush
translation and does not mention AI. Employee: Katja's manager asks her to use AI to rank staff
holiday requests. Small-business owner: Femi wants a chatbot to answer customers' questions
about refunds. Your policy must give each a clear answer.

Hint. For each card, find the line of your policy that decides it. If none does, the policy has a gap. If the same line gives the same answer to all four, the policy treats every use as the same risk.

Solution. A short first version, then the test. The first version is for a general learner, not the only form.

Example policy:

MY AI-USE POLICY      Owner: me    Version: 2    Review by: 2027-04-12
Allowed        Drafts, summaries, ideas, translation of my own text. I check every claim.
Ask first      Any text with other people's names, a client's files, money, hiring, grades,
               health. I ask the owner of the data or the rule in writing before I begin.
Not allowed    Deciding about a person's pay, job, grade, credit or health; fake sources;
               passing on private data to a tool I have not checked.
Data           Remove names and contact details before pasting. Confidential files: only
               with written permission.
Telling        I say when AI helped if a rule requires it, a person would expect to know, or
               it changes quality or risk. Customers are told they are talking to a bot.
Owner, problems  Each use has a named person; I report mistakes to them the same day.
Exceptions     Written, with an approver, an end date and a note of the outcome.

Worksheet, filled in:

Teacher Mei    Ask first: names in essays. Fix: remove names, check the school rule, read
               all comments. Result: pass after revision (v1 had no grades line).
Freelancer Karim Ask first: client files. Fix: ask the client in writing before any AI step;
               if no answer, do not use AI. Pass.
Employee Katja  Not allowed: ranking is a decision about people. Fix: Katja asks for a person
               to decide, AI at most to sort dates; escalate to the manager. Pass.
Owner Femi    Telling covers bot disclosure, but v1 had no route to a person. Fix: add a
               human contact and a refund-error report line. Pass after revision.

Pass checklist.

• Strong: permitted uses, restricted uses, disclosure rules, accountability and an exception process are all present; each of the four cards gets a clear, different answer; at least one gap is found and fixed; the policy does not assume every use has the same risk.
• Weak: one rule such as "use AI carefully" for everyone; a long policy nobody can use; no owner or review date; passing your own policy because it "seems fine" without running the four cards.

Why this works. Transparency, Explanation and Accountability: Who Answers for an AI Decision gives the one-page policy and exception process, Learning, Integrity and Disclosure: Using AI in Study and Assessment and Work, Clients and Jobs: Using AI Responsibly Where You Are Paid to Decide give the disclosure tests, and Rules, Uncertainty and Responsible Decisions: Finding What Applies to You gives the commitment you tested.

Six interview-style questions

Answer aloud or in writing in three or four sentences. These are strong short answers; yours need not match word for word, but should give a reason and one example.

Where this shows up at work

• Interviews and assessments: scenario questions ask you to name affected people, weigh options and give a decision with conditions. The worksheet is your structure.

Common mistakes

• Reading the solution first. You learn little from a solution you did not try to beat. Fix: write, then compare.
• Looking for the single right opinion. Most cards have several defensible decisions. Fix: score the quality of reasoning, evidence and conditions, not agreement with the model.
• Ticking items you only thought of. Fix: an item counts only if it is written on your sheet.
• Leaving out an owner and a date. It makes a decision look finished when nobody is accountable. Fix: always fill steps 7 and 8.

Try it yourself

1. Easy. Take your lowest-scoring challenge and redo only the step where you lost points. Did the decision change?
2. Medium. Invent a new scenario card for a friend, with all the data needed: who, what the tool does, three numbers, and what is unknown. Keep it invented. Write its pass checklist before your friend answers.
3. Hard. Test your H2 policy against a fifth card of your own (for example a student using AI for an assignment). Revise one line and record the change and the date.

Answers

Easy. Pass checklist: you redid one step only, can say which step and why, and can say whether the decision became better supported, stayed the same or got worse (worse is useful information). Weak: redoing the whole challenge without finding the gap.

Medium. Pass checklist: the card is invented and labelled; it has people, a tool, numbers that add up, and at least one unknown; the checklist states what makes an answer strong and weak and does not require one opinion; a friend could answer it from the card alone. Weak: a card that needs facts you did not print.

Hard. Pass checklist: the fifth card is specific; you found the line of your policy that decides it or noticed there was none; you changed exactly one line and wrote the date; the revised policy still gives sensible answers to the first four cards. Weak: adding a rule for every case until the policy is too long to use.

Summary

• A lab is practice with a score: write first, tick items you actually wrote, and keep your sheets.
• Good answers name affected people, give options including not using AI, say what is unknown, set conditions, and name an owner and a review date.
• Where reasonable people disagree, the checklist rewards reasons and evidence, not a particular opinion; for synthetic-media incidents, independent verification is a firm requirement.
• You can answer the six interview questions briefly: ethics versus law, equal versus fair, meaningful oversight, innovation versus risk, anonymisation, and when not to use AI.

Next: Capstone, Cheat Sheet and What's Next: Your Responsible-AI Decision File turns your sheets into a personal or workplace AI-use policy, gives you a one-page cheat sheet and a glossary, and shows where to go after this tutorial.