Task statement 6.4 · Governance, risk and responsible use

6.4 · Ethical Implications

In Wrenfield Housing, the course example, the communications team drafts a tenant FAQ for a new hardship fund with Claude. The ethical implications of AI use are its effects on the people an output reaches or describes, beyond whether its facts are right. A draft can pass every fact check and still shut some tenants out.

Why it's tested

The tempting view is that ethics is covered once the facts are checked and the AI policy is followed. A fact check asks whether the output is true, and the policy says what you must do. Neither asks who the output treats unfairly, who should know AI helped, or who answers for it. A document can be accurate and allowed and still fail the readers who need it most. The opposite mistake is treating every use of Claude as an ethical risk, so checks on brainstorms crowd out the outputs that affect people.

KnowledgeWhat ethical use asks of an AI-assisted output

Ethical questions arise when an output affects people: the tenants who read it, the applicants it describes, or the colleagues who act on it. Claude can repeat the assumptions of the human writing it learned from, and the Anthropic study recapped in 2.2 found that its answers varied with demographic details, though careful prompting reduced this. A fluent draft hides those assumptions, so the person using Claude has to ask the questions the draft cannot ask itself. This rules out any answer that treats a correct draft as finished when it reaches or describes people.

Before such an output goes out, ask these questions in order:

  1. Fairness: who could this output leave out or treat worse? Look for bias, a pattern that favours some groups over others.
  2. Respect: does it describe people in plain, accurate words, without blame?
  3. Transparency: who needs to know that AI helped, and how will they be told?
  4. Accountability: which named person signs it and answers for it?
  5. Human judgement: which decisions stay with people, and are those people still thinking for themselves?

Interactive · what the hardship fund FAQ settles as each check is done

Four review checks on the left, in the order they are done from top to bottom, each with an arrow to what it settles in the hardship fund FAQ on the right. Fairness and respect settle how tenants can apply and how they are described. Transparency settles whether the FAQ says AI helped. Accountability settles who signs it. Human judgement settles who decides each application. A check not yet done is dashed and labelled not asked, and its result is dashed and labelled at risk. REVIEW CHECK (IN ORDER, TOP TO BOTTOM) WHAT THE FAQ SETTLES Fairness and respect Who could it leave out? Does it blame anyone? NOT ASKED SETTLES LEFT OPEN SETTLED AT RISK how tenants apply, how they are described → WEB, PHONE, PAPER; HELP IN OTHER LANGUAGES → WEB FORM ONLY, ENGLISH ONLY Transparency Who needs to know AI helped? NOT ASKED SETTLES LEFT OPEN SETTLED AT RISK whether readers know AI helped → A CLOSING NOTE SAYS SO → AI'S ROLE UNSAID Accountability Who signs it and answers for it? NOT ASKED SETTLES LEFT OPEN SETTLED AT RISK who answers for the FAQ → THE COMMUNICATIONS LEAD SIGNS → NOBODY OWNS IT Human judgement Which decisions stay with people? NOT ASKED SETTLES LEFT OPEN SETTLED AT RISK who decides each application → THE PANEL DECIDES, NEVER CLAUDE → WHO DECIDES IS UNSAID Every fact in the draft was right; each check left undone is still a risk to tenants.

Step through the review of the hardship fund FAQ, from Claude's first draft to the signed version, and see what each check settles.

First draft: every fact is right, but no check is done. The FAQ offers only a web form in English, does not say AI helped, has no owner and never says who decides.

Fairness and respect checked: tenants can apply online, by phone or on paper, with help in another language, and nobody is blamed. AI's role, the owner and who decides are still open.

Transparency checked: a closing note says an AI tool helped and staff checked every answer. Nobody has signed it or said who decides.

Accountability and judgement checked: the communications lead signs, and the FAQ says a panel of three staff decides every application, never Claude.

The first row joins steps 1 and 2. Dashed and labelled AT RISK: not checked. Solid and labelled SETTLED: checked.

KnowledgeTransparency, accountability and human judgement

Transparency means being open about AI's role. Anthropic's AI Fluency course calls this Transparency Diligence: being honest about AI's role in your work with everyone who needs to know. Who needs to know depends on what readers do with the output, so a tenant applying for money has more reason to know than a colleague reading meeting notes. In one case the choice is not yours. Take advice or a decision about an individual in a high-risk area such as housing: when it is based on an AI recommendation, Anthropic's Usage Policy says that person must be told AI was used. The current version and the one starting on 12 November 2026 both require it:

Advice or a decision about one person, in a high-risk area

Required by the Usage Policy

The person must be clearly told that AI was used, and a qualified person reviews the output first, as 2.4 showed.

Your choice: none; the policy sets it.

Other outputs that reach people

Transparency Diligence

Say that AI helped where readers rely on the output or would expect to know; one line is enough.

Your choice: agreed once per kind of output.

Accountability means a named person answers for the output once it is used, like the reviewer in 2.4 who verifies and vouches for what they share. "Claude wrote it" is never the answer to a complaint, because Claude cannot explain the choice, apologise or put it right. Name the person before the work starts, because an output several people touched can end up owned by no one.

Over-reliance is accepting AI output without applying your own judgement, usually because it is fluent and has been right before. It creeps in slowly, so a reviewer who found nothing wrong in ten drafts reads the eleventh less closely. Anthropic's AI Fluency course says that even the most advanced AI systems benefit from human judgement and oversight. The firmest line is around decisions about individual people, such as who gets money, a home or a job: Claude may draft the criteria (4.1), but people judge each case.

SkillWorked example: the hardship fund FAQ

Wrenfield is opening a Winter Hardship Fund: grants of up to £250 towards heating and food, open to applications from 1 December to 31 January, each decided by a panel of three staff. The team gave Claude the fund's rules and no tenant data (policy rule 1) and asked for a twelve-question FAQ. Every fact was right, but the review found two problems:

# before (draft, Q4 and Q7)
Q4. How do I apply? Fill in the online form on our website by 31 January.
Q7. Who is it for? Tenants who have fallen behind through poor budgeting.
→ assumes every tenant reads English online; blames the reader
# after
Q4. Apply online, by phone on the income team's line, or on a paper form
    from any estate office. Call us for help in another language.
Q7. Tenants who are finding it hard to pay for heating or food this winter.
Q12. Who decides? A panel of three Wrenfield staff decides every application.
Note: We used an AI tool to help write this FAQ. Our staff checked every answer.
→ signed by the communications lead before it goes out

The cause was an assumption, not a factual error. Claude wrote for the most common reader, someone online and fluent in English, so the draft closed the fund to households without internet access or with little English. A fact check would never find this, because nothing in the draft was false. The team added other ways to apply, rewrote Q7 to describe a situation rather than blame a reader, and added Q12 so tenants know that people decide. The FAQ is general information, not advice to one tenant, so the AI note was the team's choice under Transparency Diligence. The communications lead signed the FAQ, as policy rule 2 requires, and the panel decides every application without Claude.

Ask one question: does this output reach, describe or decide something about people? Yes → check who it leaves out and how it describes them, say that AI helped where readers rely on it, name the person who signs it, and keep every decision about a person with people. No → the usual level of checking from 2.4 is enough.
Takeaway
  • A draft can be accurate and allowed and still unfair, because it can assume a typical reader and miss everyone else.
  • Be open about AI's role with everyone who needs to know; for high-risk advice or decisions about one person, the Usage Policy requires it.
  • Over-reliance is the quiet risk: a named person answers for each output, and decisions about people stay with people.

check yourself

0 / 3 answered

A town's community fund gives small grants to local groups, and a panel of four volunteers decides every application. To save time, Claude now adds a one-line recommendation to the top of each file. Over two rounds the panel agreed with every recommendation, usually within a minute, and the fund manager is concerned.Which change best keeps the decision with the panel while still saving it time?

Select one.

A sports club used Claude to draft its volunteer-coach advert and posted it after checking the dates and contact details. Last year's advert drew 30 applications, 14 from women. This year's drew 28, 3 from women. The role, dates and job board are unchanged. The new advert asks for "a dominant, competitive leader who thrives under pressure".What is the most likely root cause?

Select one.

A company's HR team uses Claude to draft feedback letters to unsuccessful job candidates, working from the interview panel's notes as the company's data rules allow. Candidates were promised the specific reasons for the decision. The HR lead wants each letter to be fair to the candidate and to have someone who answers for it.Which TWO practices best meet both aims?

Select two.

sources

  1. Claude Academy, AI Fluency, "A closer look at Diligence": Transparency Diligence as being open, and honest, about AI's role in our work with everyone who needs to know; Diligence as taking responsibility for our AI collaborations. Checked 2026-10-15.
  2. Claude Academy, AI Fluency, "A closer look at Discernment": evaluating what AI produces; even the most advanced AI systems benefit from human judgment and oversight. Checked 2026-10-15.
  3. Anthropic, "Usage Policy" (effective 12 November 2026): housing among the high-risk areas; an individual who receives advice or is the subject of a decision based on a high-risk AI recommendation must be clearly told that AI was used; a qualified person meaningfully reviews it first. Checked 2026-10-15.
  4. Anthropic, "Usage Policy", previous version (effective 15 September 2025): housing eligibility as a high-risk use case; disclose to individuals that AI helps produce advice, decisions or recommendations presented to them. Checked 2026-10-15.
  5. Anthropic, "Evaluating and mitigating discrimination in language model decisions": positive and negative discrimination found; careful prompting reduced both. Checked 2026-10-15.
  6. Wrenfield's Winter Hardship Fund, its FAQ, the draft's Q4 and Q7 and the panel of three staff are this course's example.