What you will be able to do
- Explain why using Claude never moves accountability for the result away from you
- Recognise when the people reading or affected by your work need to know AI was involved
- Spot over-reliance, including the quick skim of a draft you send under your own name
- Use AI output in decisions about people without letting bias or delegation replace human judgment
- Decide what to do in ethical gray areas that no policy covers
Key concept
Diligence — Diligence is the AI Fluency competency of taking responsibility for how you use AI and for what it produces. You decide what to hand off before you start, and you own everything that comes back once you use it.
1.Diligence: you own the outcome, not the output
Anthropic's AI Fluency courses define fluency as using AI in ways that are efficient, effective, ethical and safe. Ethics is part of that definition, not an extra added on top. The ethical side lives mostly in two of the four competencies, Delegation and Diligence. Together they form an outer loop around every piece of AI-assisted work. Before you start, Delegation asks what is right to hand off: the right task, the right tool, the right data. Once Claude has answered, Diligence asks you to take responsibility for what came back.
The simplest form of that idea comes from the builders' course: you own the outcome, not the output. The output is the text, table or ranking Claude produced. The outcome is what happens when you use it: a colleague acts on your recommendation, a reader trusts your report, a candidate is left off a shortlist. "Claude wrote it" does not move that outcome onto the tool. The course says the excuse explains nothing and excuses nothing. That holds whether the output was a draft you polished heavily or one you sent almost unchanged.
No. Disclosure and ownership are two separate duties. Telling people AI was involved is the right thing to do, but the manager still owns the outcome and still has to check the work against their own standards before it goes out.
2.Transparency with the people who rely on your work
The AI Fluency material splits Diligence into three checks: Creation, Transparency and Deployment. Transparency means being honest about AI's role with the people who read, rely on or are affected by what you produce. Before shipping anything, the builders' course asks two questions: could this output be misread or misused, and have you been open about the part AI played?
The risk is sharpest when readers take a document as your personal expert judgment. A bulletin, a commentary or an assessment carries your name, and your name carries your credibility. If Claude wrote most of it and you only skimmed it, readers are trusting a judgment you never actually made. Two things go wrong at once. Readers are misled about where the content came from, which is an attribution and transparency problem. And they act on analysis nobody properly reviewed, which is an over-reliance problem, covered in the next section. Fixing one does not fix the other. You need to be open about AI's role and also review the work properly.
3.Over-reliance: the quick skim and skill atrophy
Over-reliance means leaning on AI output more than your review of it justifies. The K-12 educators' course calls this the Dependency tension: reliance on AI replacing thinking that people need to do themselves. The creative-work course names the long-term forms. Convergence is output drifting toward the same generic answer. Skill atrophy is losing the ability to do the work yourself. It also names a short-term form: reaching for AI before you know what you want.
In knowledge work, the typical failure is a routine output that slowly slides from reviewed to glanced at. Nothing seems to break, because the drafts read well. But the review step carried the expertise, and people downstream still assume that expertise is there. The builders' course gives a test that transfers to any work: can you explain what the output does, not just what it is supposed to do? If you could not defend a claim in the draft when challenged, you have not done your Diligence on it. The creative-work course puts the minimum plainly: check finished work against your standards.
A fintech startup wants to use Claude to help loan officers draft explanations for credit decisions before those explanations are sent to applicants. Under Anthropic's Usage Policy, what must the company do to use Claude responsibly in this high-risk context?
Correct answer: B — Have a loan officer review each explanation before sending it, and disclose to applicants that AI helped draft it
- A. Incorrect. A disclaimer alone does not satisfy the human-in-the-loop requirement; a qualified professional must still review the specific decision before it goes out.
- B. Correct. Financial services is a high-risk use case under the Usage Policy, requiring a qualified professional to review AI-assisted content before dissemination and requiring disclosure to the end user.
- C. Incorrect. Avoiding disclosure does not meet the policy's transparency requirement, and internal-only use does not remove the review obligation once content reaches applicants.
- D. Incorrect. Retroactive periodic audits are not a substitute for prior review of each decision by a qualified professional before it is disseminated.
4.Bias and decisions about people
The stakes rise when AI output feeds decisions about people, such as who gets shortlisted, promoted or evaluated. The educators' course lists Bias as a core ethical tension: AI can reinforce narrow perspectives, stereotypes or existing inequities. A ranking or summary can look neutral and still carry those patterns, so a person working from it inherits them unless they check.
The course's answer is a commitment about delegation. It asks you to name the decisions you will not hand to AI and gives final judgment, relational feedback and high-stakes evaluation as examples. In practice, Claude can help organise, summarise or surface patterns, but a person makes the call. That person has to engage with the evidence directly, not just approve the AI's ordering. The builders' course adds a question about who is left out: access is a design decision, so check who your assumptions exclude. For decisions about people, that means asking whose record the summary may have misread, and whether the criteria disadvantage some groups.
| Tension | The question to ask |
|---|---|
| Dependency | When might reliance on AI replace thinking people need to do themselves? |
| Bias | When might AI reinforce narrow perspectives, stereotypes, or existing inequities? |
| Agency | Who controls how AI is used: the individual, the manager, or the system? |
| Equity of access | Who benefits from AI, and who may be excluded due to access or resources? |
| De-socialization | When could AI displace human interaction, collaboration, or relationship-building? |
| Environmental impact | How does the energy use, compute demand, or scale of AI systems factor into your ethical decisions? |
A consumer-facing startup builds a chatbot on Claude for its e-commerce site and wants to follow Anthropic's guidance on transparency. What should the company implement at the start of a customer's chat session?
Correct answer: C — A clear notice at the start of the session disclosing that the customer is talking to an AI system
- A. Incorrect. Instructing the model to impersonate a human and deny being AI directly violates the Usage Policy's prohibition on deceptive impersonation.
- B. Incorrect. A buried terms-of-service mention does not meet the requirement for clear, session-level disclosure that a customer would actually see.
- C. Correct. Anthropic's guidance for consumer-facing chatbot applications requires disclosing AI involvement at the start of the session so users know they are talking to an AI.
- D. Incorrect. Consumer chatbots are explicitly called out as needing AI disclosure at the start of the session, not exempted from it.
5.Judgment in the gray areas, and keeping it current
Organisational policy sets a minimum, but it cannot foresee every case. The educators' course is direct about this: responsible AI use is not only about following rules. It is about ethical judgment in the gray areas where policy gives no clear answer. So an action can comply with policy and still be ethically weak. Sending an unreviewed AI draft as your own analysis probably breaks no written rule.
The course suggests deciding ahead of time what you will do when you are unsure. Its examples are to pause, consult a colleague, review the relevant guidance, or pilot cautiously. The companion lesson on personal AI values says stated values help you hold your non-negotiables. One of its suggested habits is testing new tools only in low-risk contexts first. None of this is a one-off exercise. Ethical AI use is continuous, and your commitments should be revisited as your context, your colleagues and the tools change.
An HR team wants to use Claude to help draft candidate rejection emails after final hiring decisions have already been made by human recruiters. Why does this still count as a high-risk use case under Anthropic's Usage Policy?
Correct answer: D — Because employment is a designated high-risk category, so review and disclosure still apply here
- A. Incorrect. Rejection emails are not legally binding contracts, and notarization is not part of the Usage Policy's requirements.
- B. Incorrect. Claude can be used for employment-related text; the policy requires safeguards like review and disclosure, not an outright ban.
- C. Incorrect. Storage duration is not the factor that determines high-risk status; the use case category itself is what triggers the requirement.
- D. Correct. Employment is one of the categories the Usage Policy designates as high-risk due to public welfare implications, so human review and disclosure requirements apply even when a human made the underlying decision.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Disclosing that Claude drafted something transfers responsibility for its accuracy to the tool.Why is that wrong?
Disclosure and ownership are separate duties. Whoever uses AI output stays accountable for the outcome, and "AI wrote it" is no excuse.
2.If an AI draft reads well and past drafts were fine, a quick skim is an adequate review.Why is that wrong?
Output that reads fluently is not output that has been checked. Over-reliance happens when review becomes a glance while others still trust your expertise. Check finished work against your own standards.
3.When a human panel makes the final decision, bias in an AI-generated ranking of people no longer matters.Why is that wrong?
A ranking can carry bias into the panel's view. The human decision-maker has to examine the evidence and keep final judgment and high-stakes evaluation, not just approve the AI's order.
Covered in Bias and decisions about people
4.If an AI use complies with organisational policy, it is ethically sound.Why is that wrong?
Policy sets a minimum. Many ethical questions sit in gray areas that policy does not address, and those need your own judgment.
Covered in Judgment in the gray areas, and keeping it current
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“You own the outcome, not the output.”
↩︎ Diligence: you own the outcome, not the output“explains nothing and excuses nothing.”
↩︎ Diligence: you own the outcome, not the output“Could this output be misread or misused?”
↩︎ Transparency with the people who rely on your work“Can you explain what your code does, not just what it should do?”
↩︎ Over-reliance: the quick skim and skill atrophy“Access is a design decision. Check who your assumptions exclude before you call something shipped.”
↩︎ Bias and decisions about people“You own the outcome, not the output.”
↩︎ Exam trap 1 - 2.https://academy.claude.com/courses/ai-fluency-for-k-12-educators/ethics-responsible-useOfficial docs
“Afterward you take responsibility for what comes back.”
↩︎ Diligence: you own the outcome, not the output“Dependency: When might reliance on AI replace thinking students need to do themselves?”
↩︎ Over-reliance: the quick skim and skill atrophy“To protect against that concern, I will not delegate decisions related to: (e.g., final judgment, relational feedback, high-stakes evaluation)”
↩︎ Bias and decisions about people“Bias: When might AI reinforce narrow perspectives, stereotypes, or existing inequities?”
↩︎ Bias and decisions about people“Using AI responsibly isn't just about following rules; it's about ethical judgment in the gray areas where policy gives no clear answer.”
↩︎ Judgment in the gray areas, and keeping it current“When I'm unsure whether an AI use is ethical, I will: (e.g., pause, consult a colleague, review district guidance, pilot cautiously)”
↩︎ Judgment in the gray areas, and keeping it current“Ethical AI use is continuous. Revisit your commitments as students, context, and tools change.”
↩︎ Judgment in the gray areas, and keeping it current“Bias: When might AI reinforce narrow perspectives, stereotypes, or existing inequities?”
↩︎ Exam trap 3“it's about ethical judgment in the gray areas where policy gives no clear answer.”
↩︎ Exam trap 4 - 3.https://academy.claude.com/courses/ai-fluency-for-nonprofits/welcome-to-ai-fluency-for-nonprofitsOfficial docs
“AI Fluency is the ability to use AI in ways that are efficient, effective, ethical, and safe.”
↩︎ Diligence: you own the outcome, not the output - 4.https://academy.claude.com/courses/ai-fluency-for-creative-work/delegation-and-diligenceOfficial docs
“Creation, Transparency, and Deployment Diligence help you delegate responsibly.”
↩︎ Transparency with the people who rely on your work“Convergence and skill atrophy are the long-term pitfalls; reaching for AI when you do not know what you want is the immediate one.”
↩︎ Over-reliance: the quick skim and skill atrophy“Check finished work against your standards.”
↩︎ Exam trap 2 - 5.https://academy.claude.com/courses/ai-fluency-for-k-12-educators/ai-s-constitution-pedagogy-and-youOfficial docs
“Stated values help you hold your non-negotiables.”
↩︎ Judgment in the gray areas, and keeping it current
Also cited
“Diligence means taking responsibility for how you use AI.”
↩︎ Key concept