Rubric
Contents — domains, guide and mocks

Ethical implications of AI use

CCAO-F 6.411 min read · checked 21 September 2026

Task statementUnderstand the ethical implications of AI usage

Five ways an output can do harm

One outputWho could this hurt, and how badly?
  • PhysicalBodily harm — unsafe instructions, wrong clinical detail
  • PsychologicalDistress, humiliation, harassment
  • EconomicLost money, a job, credit, a claim
  • SocietalHarm to institutions, elections, public trust
  • AutonomyDeciding for someone, or deceiving them
Anthropic’s Unified Harm Framework considers these five dimensions, along with how likely and how widespread a misuse would be. Borrow it as a checklist for your own work.

Bias, and the only test most people can actually run

Bias in this context is not a moral accusation about the model; it is an observable difference in how two comparable inputs are treated. Anthropic tests for it directly before release. For political bias, the method is to give the model prompts expressing opposing viewpoints and compare the two responses, scoring them for factuality, comprehensiveness, equivalency and consistency. Responses on subjects such as healthcare and jobs are also assessed for bias relating to identity attributes.

That method is the useful part, because you can run a miniature version of it yourself in two minutes. Take the task you are about to automate, prepare two inputs that differ in exactly one respect that should not matter — the name, the pronoun, the school, the postcode, the stated political view — and compare the outputs. If they differ in substance rather than in the detail that changed, you have found something. If they differ in tone, warmth or hedging, you have found something subtler and probably more consequential, because tone is what a reader acts on.

Two prompts, one difference

Invites an impressiontext

Here is Sarah's
self-assessment.

Write her performance
review. Say whether
she's ready for
promotion to team
lead.

Fixes the criteriatext

Here is a self-
assessment and our
four promotion criteria.

For each criterion,
quote the evidence in
the document and mark
it met, partly met or
not evidenced.

Do not recommend a
decision. Use "the
employee" throughout.
The weak prompt asks for an impression, and impressions are where stereotype leaks in. The strong one fixes the criteria and demands evidence, which makes a swap test meaningful.

Honesty: disclosure, attribution and impersonation

The Usage Policy draws a hard line at impersonating a human being, and it requires consumer-facing chatbots to tell people they are talking to an AI. In high-risk domains it asks for disclosure of AI involvement to the people affected, at minimum at the start of each session. Behind all three is one principle: a person is entitled to know what they are dealing with, because that knowledge changes how much weight they give it and what they do next.

Most workplace dishonesty with these tools is quieter than impersonation. It is a condolence note signed by a manager who never read it. A personal-sounding donor letter from a named trustee, mass-produced. A consultation response that claims to be one organisation’s considered view and is in fact a first draft nobody edited. None of these breaks a rule that anyone will enforce. All of them trade on a reader’s assumption about how much human attention they received, and that assumption is precisely what is being spent.

Two different questions

Compliance asks

  • Is it permitted by the Usage Policy?
  • Does our AI policy allow it?
  • Is there a lawful basis for the data?
  • Has the required reviewer signed off?

Ethics asks

  • Who is affected, and are they in the room?
  • Would they object if they knew how this was made?
  • Is one group treated differently for no good reason?
  • If it is wrong, who carries the consequence?
A task can clear everything on the left and still fail on the right. The exam’s ethics items live in the right column.

Accountability does not transfer

Anthropic invests heavily on its side of the line. It develops policy with external domain experts through policy vulnerability testing — during elections, for example, working with the Institute for Strategic Dialogue, which led to informational banners pointing users to authoritative voting resources. It runs fine-tuned classifiers that detect violations in real time, monitors aggregate patterns through privacy-preserving techniques, and evaluates models for bias and for cyber and weapons risk before deployment.

None of that makes your output correct. The safeguards address harm at population scale; they cannot know whether this clause reference in this letter to this policyholder is real. Anthropic’s own help documentation says plainly that Claude can hallucinate, can fabricate authoritative-sounding quotations, and should not be relied on for high-stakes advice without scrutiny — and its prompt-engineering guidance closes with the same caution, that techniques for reducing hallucination significantly reduce but do not eliminate it. Whoever signs the work owns the work.

There is a legal echo of this in the GDPR, which gives people the right not to be subject to a decision based solely on automated processing where it produces legal effects or similarly significantly affects them. The law, the Usage Policy and ordinary professional ethics converge on one sentence: a person decides, and that person is answerable.

An output checked for the absent person

  • Passes: I can name who is affected by this outputThe policyholder whose claim was declined
  • Check: They will be told AI was involvedDisclosure line drafted but not yet in the template
  • Passes: A swap test found no unexplained differenceNames and postcodes varied; output unchanged
  • Fails: Nothing implies more human attention than it receivedSigned “I have personally reviewed your case”
  • Passes: A named person is answerable for the decisionThe adjuster, not the drafting tool
  • Missing: Someone checked the figures and referencesTwo clause numbers still unverified
Run this after the accuracy check, not instead of it. Accuracy is 2.1; this asks a different question about the same draft.

Traps the wrong answers are built from

Tempting but wrongDo this instead
Treating “it is permitted” as the end of the questionAsk separately who is affected and whether they would object if they knew how the output was made.
Asking for an impression of a personFix the criteria, anonymise the input and ask for quoted evidence against each criterion.
Assuming bias would be obvious if it were thereRun a swap test — change one irrelevant attribute and compare the two outputs.
Letting a reader believe a draft had human attention it did notDisclose AI involvement, and make sure whoever signs has actually read it.
Citing the model as the reason for a decisionName the person who decided; accountability does not transfer to a tool.

You should now be able to

  • Name the five harm dimensions and use them to assess a proposed use.
  • Run a swap test on a prompt that makes judgements about people.
  • Rewrite an impression-seeking prompt into an evidence-and-criteria prompt.
  • Explain why disclosure matters even where no rule requires it.
  • Separate the compliance question from the ethical question in a scenario.

Practice questions

Original questions written for this lesson, in the exam’s style. Answer first, then open the reasoning — every option is explained, including why the wrong ones are tempting.

  1. Question 1

    A bereavement team at a utility company uses Claude to draft condolence letters to the families of deceased account holders. The letters are warm, accurate and signed by the customer care director, who does not read them individually.

    What is the strongest ethical objection?

    1. AThe letters process personal data about a deceased person without consent.
    2. BThe signature implies personal attention the recipient did not receive.
    3. CThe Usage Policy prohibits AI-generated correspondence to consumers.
    4. DClaude cannot write with genuine emotion, so the letters will read as hollow.
    Show answer and reasoning
    1. AIncorrect. It raises a data question worth asking, but it is not what makes the practice feel wrong to the recipient.
    2. BCorrect. The letter trades on an assumption about human attention, which is exactly the resource being spent without the reader knowing.
    3. CIncorrect. No such prohibition exists; media and consumer-facing content carry conditions rather than a ban.
    4. DIncorrect. The scenario stipulates the letters are warm, and quality is a separate question from honesty.

  2. Question 2

    A lettings agency asks Claude to summarise each tenancy application and say whether the applicant “seems like a reliable tenant”. An analyst suspects the summaries are not treating all applicants alike.

    What is the most effective way to find out, and the best remedy?

    1. AAsk Claude whether its own summaries are biased, and then act on the answer it gives.
    2. BSwitch to a larger model, since stronger models exhibit less bias.
    3. CRun two applications differing only in name, then ask for evidence against fixed criteria.
    4. DAdd “be unbiased and fair to all applicants” to the prompt.
    Show answer and reasoning
    1. AIncorrect. A model's self-report about its own outputs is not evidence and cannot substitute for comparing actual outputs.
    2. BIncorrect. Model capability is not the lever here; the prompt is asking for an impression rather than for evidence.
    3. CCorrect. The swap test isolates the irrelevant attribute, and moving from impression to criteria removes the space where stereotype operates.
    4. DIncorrect. An instruction to be fair does not change what the request is asking for, and leaves nothing that can be checked.

  3. Question 3

    A local council plans a website assistant that answers residents’ questions about bin collections, council tax bands and planning applications, drawing on published council pages.

    Which two obligations follow most directly from the ethical and policy considerations in this lesson? (Select 2.)

    1. AResidents must be told they are interacting with an AI rather than a person.
    2. BA named officer must remain answerable for what the assistant tells residents.
    3. CEvery resident must give consent before their question is processed.
    4. DThe assistant must refuse all planning questions as a high-risk legal use.
    5. EThe council must publish the model’s training data before launch.
    Show answer and reasoning
    1. ACorrect. Consumer-facing chatbots are expected to disclose that the user is talking to an AI, and residents act differently when they know.
    2. BCorrect. Accountability does not transfer to a tool; someone in the council owns the answers it gives.
    3. CIncorrect. No such requirement follows; the relevant obligation is disclosure of AI involvement, not individual consent.
    4. DIncorrect. Signposting published planning information is not legal advice, and a blanket refusal would remove legitimate public value.
    5. EIncorrect. This is not an obligation any source here establishes, and it would not help a resident judge an answer.

  4. Question 4

    Which statement best describes the relationship between Anthropic’s safeguards and a user’s responsibility?

    1. ASafeguards handle harm at scale; the user still owns the accuracy and fairness of their own output.
    2. BBecause Claude is evaluated for bias before release, user-side bias checks are unnecessary.
    3. CReal-time classifiers block any output that would harm someone, so review is a formality.
    4. DResponsibility rests with Anthropic, since it controls the model’s behaviour.
    Show answer and reasoning
    1. ACorrect. Classifiers, bias evaluations and policy testing address population-level risk and cannot verify one specific claim in one document.
    2. BIncorrect. Pre-release evaluation reduces systematic bias but says nothing about how a particular prompt shapes a particular output.
    3. CIncorrect. Classifiers detect policy violations, not factual errors or unfairness in ordinary business writing.
    4. DIncorrect. The organisation deploying an output to affected people is the one accountable for it, in policy and in law.

Practise every question in Domain 6 →

Sources

Drafted with AI assistance and checked against the sources above; expert review is in progress. Spotted an error? Tell us and it gets fixed, dated and listed on how this is written.