Home › Study guides › CCAO-F › Domain 6 › Lesson 6.4

CCAO-F · Domain 6 · 15% of the exam · Lesson 6.4 · 19 min read

The ethics of using AI at work

The duties that stay with you when Claude drafts work that affects people: disclosure, accountability, fairness, honesty and respect for others' work.

Written against objective 6.4 of the official CCAO-F exam guide (Version 1.0, effective July 2026). An independent resource, not affiliated with Anthropic; the practice questions are written from scratch.

6.4.1 Why a faster way to mark raises questions about people

Miriam teaches English at a secondary school. It's the end of term, and 120 persuasive essays from four Year 10 classes are waiting on her desk. Written feedback usually costs her three weekends, and she also owes the classes a practice test before the exam. A colleague shows her how Claude can draft comments against the class rubric in seconds, and she decides to use it for both jobs.

Within the first hour, questions start arriving. Should students and parents be told? If Claude suggests a mark, whose mark is it? About fifteen of her students are still learning English: will their feedback be as good as everyone else's? Claude's draft of a letter to parents quotes "a Year 10 student" who loves the new feedback, but who said that? Which student information should go into Claude, and will she still read every essay?

None of these is a question about what Claude can do. It can draft good comments. Every question is about the people around the work: students, parents, colleagues, the authors whose texts end up in the test. Claude can't answer them for her, because it knows only what it's given, not what those people expect or what she owes them. These are the ethical implications of using AI: the effects your use has on other people, and the duties that come with them.

6.4.2 Diligence: the responsibilities that stay with you

It's tempting to hope that a carefully built AI takes care of the ethics for you. Resist that hope. Anthropic's AI Fluency course has a name for the competency that keeps these duties with you: Diligence, taking responsibility for your collaborations with AI. In practice, that means answering for what you create with AI and how you share it. The course's lesson on Diligence, which aims to help you "understand the ethical implications of AI collaboration", splits that responsibility into three parts.

The course adds that personal, academic and professional contexts carry different expectations for disclosure and verification. A shopping list drafted with Claude owes nobody an explanation. Feedback that shapes how a fifteen-year-old sees her own writing does.

Think of a hire car. The rental company keeps it roadworthy, but the driver answers for where it goes, who rides in it and how it's driven. Put precisely: Anthropic sets the rules for how Claude may be used, and you answer for the use you make of it and for the people that use affects.

6.4.3 Tell people when AI shaped something that matters to them

Here is the question that trips people up: if Miriam reads and edits every comment, does she still need to mention Claude? Two opposite mistakes are common: disclosing only when AI did all the work, or never, because how feedback gets written is nobody's business. Both miss the real test. Would the people receiving it read it, rely on it or feel about it differently if they knew? Students take written feedback as their teacher's own reading of their work. Anthropic's research on how university teachers use Claude notes that students have begun raising concerns about their professors' AI use.

Norms also differ by field: journalism, academic publishing, healthcare and law each have their own, and your school or employer may have written rules. Anthropic's Usage Policy marks the outer edge. It forbids impersonating a human by presenting AI output as human-generated, and it requires any consumer-facing chatbot to tell people they are talking to AI. Inside that edge, the test decides: the more an output shapes how someone is judged or treated, the stronger the case for telling them.

A useful disclosure is specific: what AI did, what the person did, who is responsible and how to question the result. The AI Fluency course calls a note like this a diligence statement: an open acknowledgment of AI's role, with your commitment to responsibility for the final output. Here is Miriam's, printed at the top of each feedback sheet. Look at the lines that start "What Claude did" and "What I did": they separate Claude's part from hers.

How your feedback was written

This term I used an AI assistant, Claude, to help draft the written comments on your persuasive essays.

What Claude did: it suggested comments against our class rubric. It never saw your name, and it did not give you a mark.

What I did: I read every essay myself, rewrote or deleted comments that did not fit, and decided every mark.

The comments and the marks are my responsibility. If a comment doesn't make sense to you, or you think I missed something in your essay, come and talk to me.

6.4.4 You own the mark, so keep the judgment to own it

Now the belief that does the most damage: once Claude has suggested a mark, part of the responsibility has moved to Claude. It hasn't. The person who uses an output owns it. When a parent asks why their son got a C, "Claude suggested it" is not an answer Miriam can give. One professor in Anthropic's education research put the duty bluntly: students pay for the teacher's time, not the AI's.

Vouching has a condition people forget: you can only vouch for what you can judge. If Miriam stops reading essays and only skims Claude's comments, three things slip. She can no longer tell a sharp comment from a generic one. She loses what marking teaches her about her classes, which is where next term's lessons come from. And the skill itself fades with disuse, the way a language does when you stop speaking it. That is deskilling: expertise worn away because a tool now does the part of the work that kept it sharp.

The way out is to use Claude to extend her judgment, not replace it. Another professor in the same research put it neatly: use AI as "a thought partner, not a thought substitute". So Miriam reads every essay first and jots down what she sees. Claude then turns her notes and the rubric into drafted comments.

Over-reliance has a quieter cost too: the people whose work changes. Daniel, the department's teaching assistant, supports the students still learning English and reads their drafts with them. A plan that sends all feedback through Claude could sideline him without anyone deciding it should. Miriam asks him to review the feedback for the students he supports, which is exactly where his judgment is worth most.

6.4.5 Fair to every student it affects

Is the feedback as good for the fifteen students still learning English as for everyone else? That is fairness: an output that affects people should give each of them comparable quality and treatment. The difficulty is that unfairness rarely announces itself. Anthropic's own explanation of AI bias notes that it can be as quiet as giving better-quality answers in some languages than in others.

A tilt can enter from two directions. One is the data: AI models learn from huge amounts of human text, and patterns in that text can lean one way. The other is closer to home: your framing and your material. Suppose Miriam's rubric rewarded "sophisticated, native-like expression". Claude would apply it faithfully, and a strong argument from a student still learning English would get feedback about grammar instead of ideas. Bias in the data or the framing becomes bias in the result, applied to every student at speed.

So Miriam writes fairness into the brief. Look at the third line: it keeps language errors inside their own criterion instead of letting them colour the rest.

Help me draft feedback on Year 10 persuasive essays. I will read every essay and decide every mark myself.

Rubric, in this order: 1) the argument and its reasons, 2) use of evidence, 3) organisation, 4) accuracy of language.

Some students are still learning English. Judge their argument, evidence and organisation exactly as you would anyone else's; language is assessed only under criterion 4.

Give every essay the same depth: one strength and one next step for each criterion, at about the same length.

For language, name up to three patterns to practise, not a list of every error.

A brief can't prove its own fairness, so Miriam checks the results in pairs. Anthropic tests Claude for political bias the same way: it asks about one topic from two opposing sides and checks that both answers get the same depth and effort. Miriam takes five pairs of essays with arguments of similar strength, one from a fluent writer and one from a student still learning English. She compares the feedback side by side for length, tone, attention to ideas and a concrete next step.

The practice test gets a check too. Anthropic's AI Fluency course for educators lists "check for unintended bias" among the final checks on a quiz. A reading passage that assumes knowledge of one country's sport tests background, not reading.

6.4.6 Nothing invented, nothing taken without asking

Back to that letter to parents. Claude's draft ends with a warm touch: "'The new feedback finally showed me what to fix.' Year 10 student." No student said it. Claude wrote what a happy student might plausibly say, and the letter would present that invented voice as real. This is honesty: no fabricated testimonials, fake endorsements or quotes put in real people's mouths. Anthropic's Usage Policy prohibits generating fake reviews and comments, and creating fake personas that mislead people about where content came from. The working rule: treat every story, quote and endorsement as unverified until you've tied it to a real source and confirmed you may use it.

The practice test raises the other half: respect for other people's work. The Usage Policy also rules out infringing intellectual property, and plagiarising or submitting AI-assisted work without proper permission or attribution. Here is what Miriam finds in her drafts, and what she does about each.

Consent and respect apply to the students' data as well as their words. They wrote their essays for their teacher, not for an AI tool, so Miriam shares only what the feedback needs: the essay text, without names. Exactly which personal data may go in is a matter for data rules and school policy. She also turns down a tempting shortcut: uploading every student's learning-support plan to a Claude Project "so feedback is personalised". Claude draws on a Project's knowledge in every chat inside it, so those records would sit behind every conversation when the feedback never needed them.

6.4.7 The exam traps

Every trap here either lets the output stand because it looks fine or throws the work away. The right answer keeps the work and adds the missing safeguard.

• ✗ Accepting Claude's suggested mark or decision because it looks consistent. ✓ Read the work and decide yourself. You own what you use, and consistency is not correctness.
• ✗ Staying quiet about AI's role because you edited everything, or adding "AI may have been used". ✓ Tell the people affected what AI did, what you did and who is responsible.
• ✗ Assuming the output is fair because everyone got the same prompt. ✓ Fix any framing that tilts, then compare outputs for similar cases from different groups.
• ✗ Keeping a story, quote or borrowed passage because it sounds right. ✓ Tie it to a real source, confirm you may use it and credit its author; otherwise remove it.
• ✗ Giving Claude everything about the people involved "so it can personalise". ✓ Share only what the task needs, and keep sensitive records out of Projects that don't need them.
• ✗ Banning AI from the task after an ethical slip. ✓ Keep the use and add the missing safeguard: disclosure, review, a fairness check or verification.

6.4.8 Put it together: run a fairness check on AI-drafted feedback

You now have every piece: Diligence as the frame, then transparency, accountability backed by your own judgment, fairness, honesty and respect. The exercise practises the fairness check, because unequal treatment is the one problem you can't see by rereading a single output.

These habits carry into Domain 7. Diagnosing a poor output (7.1) often starts where this exercise did, with the framing in the brief. Adjusting your approach based on feedback (7.2) includes feedback from the people an output affects. And optimising a workflow (7.3) should make disclosure, review and fairness checks faster, never quietly remove them.

Key takeaways

• ✓ Diligence is taking responsibility for your collaborations with AI, in three parts: creation, transparency and deployment.
• ✓ Tell people when AI shaped something that matters to them, saying what AI did, what you did and who is responsible.
• ✓ The person who uses an output owns it, so keep doing the work that lets you judge it, and consider how the change affects colleagues' roles.
• ✓ Bias in the data or your framing becomes bias in the result; write fairness into the brief and compare outputs for similar cases side by side.
• ✓ Treat every story, quote and endorsement as unverified until it's tied to a real source and permission to use it.
• ✓ Respect others' work and data: credit and have the right to use what you borrow, and give Claude only what the task needs.
• ✓ The right response to an ethical risk keeps the work and adds the safeguard; it neither trusts the output as it is nor bans the tool.

Check your understanding

4 questions written for this lesson, then one from the CCAO-F question bank on the same topic. Every answer option is explained, including the ones you did not pick. Nothing is stored.

Sources

• A closer look at Diligence (AI Fluency)
• Usage Policy
• How educators use Claude (Anthropic Education Report)
• Applying AI Fluency to learning materials and assignments (AI Fluency for educators)
• Why does bias exist in AI models?
• What are projects?
• The CCAO-F exam guide, summarised