# D6 · Governance, Risk, and Responsible Use Appropriate use cases, data classification and privacy law, organisational AI policy, shadow AI, IP and disclosure, bias and accountability, Anthropic usage policies, enterprise controls, and escalation paths. import { Accordions, AccordionItem, Tabs, TabItem } from '@prosefly/astro-components'; At **15%** this is one of the heaviest domains – roughly **9 of 60 items**. It tests whether you use Claude *responsibly*: choosing appropriate use cases, classifying and protecting data, respecting privacy law and organisational policy, avoiding shadow AI, handling IP and disclosure honestly, mitigating bias, and remembering that **a human always owns the output**. Many items hinge on a single instinct: *should this data or this decision be here at all?* ## Learning objectives By the end of this page you should be able to: 1. Distinguish **appropriate vs. inappropriate** use cases. 2. Apply **data classification** (public/internal/confidential/restricted) and recognise **PII, PHI, PCI**. 3. Account for **regulatory/privacy** considerations (GDPR, HIPAA, sector rules). 4. Comply with **organisational AI policy** and acceptable-use, and avoid **shadow AI**. 5. Handle **IP/copyright** and **disclosure/transparency** of AI use. 6. Address **bias/fairness** and hold to **human accountability** for outputs. 7. Describe **Anthropic's usage policies** at a high level and **enterprise controls** (SSO, audit logs, retention, no-training-on-data for commercial plans). 8. Know the **escalation paths** for risk and compliance questions. --- ## 6.1 Appropriate vs. inappropriate use | Appropriate | Inappropriate (without safeguards) | | --- | --- | | Drafting, summarising, analysing, brainstorming | Making final regulated decisions (medical, legal, hiring, credit) alone | | Grounded Q&A over documents you may use | Feeding restricted data into unapproved tools | | Research synthesis with verification | Presenting unverified output as authoritative fact | | Structuring and reformatting information | Generating deceptive, harmful or infringing content | The judgment is not "AI good/bad" but "**is this the right task, with the right data, under the right safeguards, with a human owning the outcome?**" :::tip[Exam signal] Inappropriate-use options usually combine a *sensitive data type* or *regulated decision* with *no safeguard, no review, or an unapproved tool*. Appropriate-use options keep humans deciding and route sensitive data through approved channels. ::: ## 6.2 Data classification Know your data before you paste it. | Class | Definition | Handling with Claude | | --- | --- | --- | | **Public** | Already public | Free to use | | **Internal** | Non-public, low sensitivity | Use on approved tools per policy | | **Confidential** | Sensitive business data | Only on approved/enterprise tooling; minimise | | **Restricted** | Highly sensitive/regulated | Usually prohibited or tightly controlled; check policy first | Special categories that raise the stakes regardless of class: | Acronym | Meaning | Regime | | --- | --- | --- | | **PII** | Personally identifiable information | GDPR, privacy laws | | **PHI** | Protected health information | HIPAA | | **PCI** | Payment card data | PCI DSS | :::caution[Minimise and check first] Before pasting anything sensitive, ask: *what class is this, does policy allow it on this tool, and can I minimise or redact it?* When unsure, treat it as more sensitive and check policy or escalate – do not guess. ::: ## 6.3 Regulatory and privacy considerations | Regime | Applies to | Implication for Claude use | | --- | --- | --- | | **GDPR** | EU personal data | Lawful basis, minimisation, data-subject rights; be careful pasting PII | | **HIPAA** | US health data (PHI) | Requires appropriate safeguards/agreements; do not paste PHI into unapproved tools | | **Sector rules** | Finance, legal, education, etc. | Confidentiality, record-keeping, professional-conduct duties | You are not expected to be a lawyer. You are expected to **recognise when a regime is in play** and route the decision to the right owner (privacy/legal/compliance) rather than proceed on your own judgment. ## 6.4 Organisational AI policy, acceptable use, and shadow AI - **AI policy / acceptable use** defines which tools are approved, what data may be used, when review is required, and disclosure rules. Compliance is not optional convenience. - **Shadow AI** is using unapproved AI tools (or approved tools with disallowed data) outside governance – a major risk because it bypasses controls, retention guarantees and audit. :::danger[Shadow AI] Pasting confidential company data into a personal, unapproved AI account to "get work done faster" is shadow AI. It is a governance violation even if the output is good, because the data has left the organisation's controls. The correct move is to use the **approved** tool or escalate for approval. ::: ## 6.5 IP, copyright, and disclosure | Topic | Principle | | --- | --- | | **Input IP** | Do not feed third-party copyrighted or licensed material into tools in violation of its licence | | **Output IP** | Review generated content for infringement before external use; you are accountable for what you publish | | **Attribution** | Verify and attribute sources; do not present others' work as original | | **Disclosure/transparency** | Follow policy and context norms on disclosing AI involvement (e.g., in academic, legal, or customer contexts) | Transparency is situational but the safe default the exam rewards is **honesty**: disclose AI use where policy, professional norms, or the audience's reasonable expectations require it, and never misrepresent AI output as human-authored where that matters. ## 6.6 Bias, fairness, and accountability - **Bias** can appear in outputs (skewed framing, stereotyped assumptions). Mitigate with neutral prompting, symmetric structures, representative examples, and human review (see D2). - **Fairness** matters most where outputs affect people – hiring, credit, benefits, discipline. These need human review for biased or discriminatory language and reasoning. - **Accountability** is the anchor principle: **the human owns the output.** Claude assists; a person is responsible for what is decided, sent, or published. "The AI said so" is never a valid defence. :::tip[Exam signal] Any option that shifts responsibility to the model ("Claude decided", "the AI is accountable") is wrong. Correct options keep a named human accountable for the outcome. ::: ## 6.7 Anthropic's usage policies (high level) At a conceptual level, Anthropic's usage policies **prohibit** using Claude for things like: - Generating harmful, illegal, deceptive or abusive content. - Compromising privacy or security, or enabling surveillance and profiling that harms people. - Producing malware, weapons-enabling content, or content that exploits or endangers minors. You do not need the exact text; you need to recognise that these prohibited categories exist and that legitimate business use stays well inside them. When a request approaches a boundary, decline and escalate rather than improvise. ## 6.8 Enterprise controls Commercial (Team/Enterprise) plans provide the controls governance depends on. | Control | What it gives | Why it matters | | --- | --- | --- | | **SSO** | Central identity/login | Access control, off-boarding | | **Audit logs** | Record of activity | Accountability, investigations, compliance | | **Data-retention controls** | Configure how long data is kept | Privacy/regulatory requirements | | **No training on your data** | Commercial data is not used to train models | Confidentiality assurance for business use | :::note[The no-training guarantee] On commercial plans, your business data is **not used to train models**. This is a key reason to use approved, sanctioned tooling rather than a personal account, and a frequent correct-answer detail on the exam. ::: ## 6.9 Escalation paths Knowing when to stop and route the question is itself a competency. | Trigger | Escalate to | | --- | --- | | Restricted/regulated data or an unclear data question | Privacy / security / compliance | | Legal, contractual or IP uncertainty | Legal | | A need that requires API/agent/automation | Developers / Architects (D4) | | A request approaching a usage-policy boundary | Decline, then compliance/security | | Fairness/discrimination concerns in a people decision | HR / legal / compliance | When unsure, **escalate rather than proceed**. Proceeding on a guess with sensitive data or a regulated decision is the wrong answer pattern. --- ## 6.10 A data-handling decision tree Before any sensitive data touches Claude, route it through a short, repeatable check. This turns the classification table into a decision. ```text What class is the data? (public / internal / confidential / restricted) │ ├─ Public ─► Use freely on an approved tool. │ ├─ Internal ─► Use on an APPROVED tool, per policy. │ ├─ Confidential ─► Approved/enterprise tooling only; MINIMISE; central owner. │ └─ Restricted (or any PII/PHI/PCI) ─► Is it explicitly permitted by policy on this specific approved tool? │ ├─ Yes ─► Minimise/redact to the least needed; proceed under controls. │ └─ No / Unsure ─► STOP. Do not paste. Escalate to privacy/security/ compliance. Default to treating it as more sensitive. ``` | Signal in the stem | Correct instinct | | --- | --- | | "personal free/unapproved account" | Shadow AI – use the approved tool or escalate | | "PHI / patient records" | HIPAA + policy question first; minimise; escalate if unclear | | "payment card numbers" | PCI DSS – do not paste into an unapproved tool | | "not sure if it's allowed" | Default to caution; escalate; do not guess | | "delete the chat afterward" | Deletion does not undo exposure or a policy breach | :::tip[Exam signal] The safe default is **caution + escalation** when a data question is unclear. Options that "proceed and delete afterward", "ask Claude if it's allowed", or "assume it's fine" are wrong; Claude is not an authoritative policy source. ::: ## 6.11 Accountability, transparency and IP in practice Three responsible-use ideas recur as distractor traps. Nail the exact principle for each. | Principle | The rule | The trap it defeats | | --- | --- | --- | | **Human accountability** | A named human owns every output that is sent, published or decided | "The AI decided / the AI is accountable" | | **Transparency / disclosure** | Disclose AI use where policy, professional norms or audience expectations require it | "Never disclose" / "disclosure is optional here" | | **IP & copyright** | Do not feed material in violation of its licence; review generated output for infringement before external use | "Generated content is automatically original and safe" | **Before/after – handling an error blamed on "the AI":** ```text Before: "The AI wrote it, so the AI is responsible for the mistake." Reality: The person who used and sent the output is accountable. Claude assists; it does not absorb responsibility. After: "I own this output. I should have verified the figure before sending; here's the correction and the review step we're adding." ``` :::caution[Disclosure is situational, honesty is not] Whether you must *disclose* AI assistance depends on policy and context, but you must never *misrepresent* AI output as human-authored where that matters, and never present unverified output as authoritative fact. When policy requires disclosure, omitting it is a violation regardless of intent. ::: ## 6.12 Common misconceptions | Misconception | Reality | Why it matters on the exam | | --- | --- | --- | | "Deleting the chat undoes the exposure." | The data already left the org's controls. | Shadow-AI item. | | "A personal free account is fine if the output is good." | Output quality does not cure a governance breach. | Shadow-AI distractor. | | "'The AI decided' explains an error." | A named human owns the output. | Accountability item. | | "Claude can tell me whether data use is compliant." | Claude is not an authoritative policy source; escalate. | Self-report/authority distractor. | | "Disclosure of AI use is always optional." | It is required where policy/norms/expectations demand it. | Transparency item. | | "Generated content is automatically original." | You are accountable for infringement you publish. | IP-review item. | | "A paid plan removes the need for human review." | No plan removes review; plans add controls, not judgement. | Control-vs-review distractor. | | "Regulated decisions are fine if Claude is confident." | Fairness-sensitive decisions stay with an accountable human. | Decision-ownership item. | ## 6.13 Scenario walkthrough – a well-meaning shortcut with sensitive data **Scenario.** Under deadline, a recruiter wants Claude to help shortlist candidates. He plans to paste 200 CVs (containing names, addresses and dates of birth) into his personal free AI account "because the work account is slow", have Claude rank and auto-reject the bottom half "to remove human bias", omit any mention of AI use in the hiring report "to look rigorous", and delete the chat afterward "so there's no trace". **Expert reasoning trace.** 1. **Spot the shadow AI.** Pasting CVs into a personal, unapproved account is shadow AI: confidential PII leaves the org's controls, retention and audit. Deletion afterward does not undo the exposure. Reject the personal account – use the approved tool or escalate. 2. **PII triggers privacy law and policy.** CVs contain personal data (GDPR-type obligations); the first question is whether policy permits this data on this tool, minimised. If unclear, escalate to privacy/compliance rather than proceed. 3. **Auto-reject is the fairness trap.** Letting Claude rank and reject candidates does not "remove bias" – it can launder bias and removes human accountability for a regulated, people-affecting decision. Hiring stays a human decision with review for biased/discriminatory language. 4. **Non-disclosure violates policy.** If policy requires disclosing AI assistance in the hiring process, omitting it "to look rigorous" is a transparency violation, not a virtue. 5. **Deletion is not a control.** Deleting the chat neither reverses the data exposure nor satisfies record-keeping obligations; it may worsen the audit position. 6. **The responsible path.** Use the approved enterprise tool, confirm policy permits the data (minimise/redact), keep humans deciding with bias review, disclose AI assistance as policy requires, and retain appropriate records. **Exam-correct decision:** stop the personal-account use (shadow AI), settle the PII policy question via escalation, keep humans accountable for shortlisting with bias review, disclose AI use per policy, and follow retention rules. **Not** the personal account, **not** auto-reject, **not** hidden AI use, **not** "delete to remove the trace". --- ## Exam traps in this domain | Trap | Why it is wrong | | --- | --- | | "Paste confidential data into a personal AI account to move faster" | Shadow AI; bypasses controls, retention and audit | | "Let Claude make the final hiring/credit/medical decision" | Regulated, fairness-sensitive decisions stay with an accountable human | | "'The AI decided' explains the error" | Accountability rests with the human owner of the output | | "It's fine to use PHI/PII anywhere since Claude is secure" | Regime + policy govern the data; use approved tooling and minimise | | "No need to disclose AI use, ever" | Disclosure follows policy, norms and audience expectations | | "Free personal plan is fine for confidential enterprise data" | No-training and admin controls live in commercial plans; use approved tooling | | "Publish generated content without an IP/accuracy check" | You are accountable for infringement and errors you publish | | "Proceed when unsure about a regulated data question" | Unclear/regulated situations should be escalated, not guessed | | "Delete the chat afterward, so the exposure is undone" | Deletion does not reverse data that already left the org's controls | | "Ask Claude whether the data use is compliant" | Claude is not an authoritative policy source; escalate to compliance | | "Auto-reject candidates to remove human bias" | This launders bias and removes accountability; humans decide with review | | "A commercial plan removes the need for human review" | Plans add controls, not judgement; review is still required | --- ## Practice questions Each item states how many responses to select. Attempt before revealing. A. Efficient; proceed since the output will be good. B. Shadow AI; use the organisation's approved, sanctioned tool instead, or escalate for approval. C. Acceptable if they delete the chat afterward. D. Fine because roadmaps are not personal data. **Answer: B.** Putting confidential data into an unapproved personal account is shadow AI – it leaves organisational controls, retention and audit, regardless of output quality. Deleting afterward (C) does not undo the exposure; roadmaps are confidential business data even if not PII (D). A. Which model is cheapest. B. Whether HIPAA safeguards and organisational policy permit PHI on the specific tool, and whether the data can be minimised — escalate to compliance if unclear. C. The output format. D. Whether the chat is fast. **Answer: B.** PHI triggers HIPAA and policy questions that must be settled before use; minimise and escalate if unclear. Model cost (A), format (C) and speed (D) are irrelevant to the governing legal/policy question. A. The model is accountable for its outputs. B. The human who used and sent the output is accountable for it. C. Accountability depends on the model tier. D. No one is accountable for AI output. **Answer: B.** Human accountability is the anchor principle: a person owns what is sent or published. 'The AI decided' is never a valid defence (A, D); tier is irrelevant (C). A. Data is not used to train models on commercial plans. B. Audit logs and SSO support access control and accountability. C. The model becomes smarter on paid plans. D. Personal plans offer better governance. E. Commercial plans remove the need for human review. **Answer: A and B.** The no-training guarantee and admin controls (audit logs, SSO) are why business data belongs on commercial tooling. Paid plans do not change model intelligence (C); personal plans lack org governance (D); and no plan removes the need for human review (E). A. Let Claude rank and reject candidates automatically to remove bias. B. Keep humans accountable for decisions and review outputs for biased or discriminatory language, following policy and law. C. Trust the model because it is neutral. D. Use the cheapest model to reduce cost. **Answer: B.** Hiring is fairness-sensitive and regulated; humans stay accountable and review for bias. Automatic ranking/rejection (A) can launder bias and removes accountability; models are not inherently neutral (C); cost (D) is beside the point. A. Assume it is fine and proceed. B. Treat it as more sensitive and escalate to privacy/security/compliance before use. C. Ask Claude whether it is allowed. D. Use it but delete the chat afterward. **Answer: B.** Uncertainty about restricted data should be resolved by escalation, defaulting to caution. Proceeding on a guess (A, D) risks a violation; asking Claude (C) is not an authoritative policy source. A. Nothing; generated content is automatically original and accurate. B. Review for accuracy and potential IP/copyright issues, since the team is accountable for what it publishes. C. Only a spell check. D. Publish first, fix later. **Answer: B.** The publisher is accountable for accuracy and IP; external content needs review before release. Generated content is not automatically safe (A); spell check (C) is insufficient; publish-first (D) is irreversible and unsafe. A. Summarising an internal meeting on an approved tool. B. Using Claude to generate deceptive content impersonating a real person to mislead customers. C. Brainstorming campaign ideas. D. Drafting a policy FAQ from an approved handbook. **Answer: B.** Deceptive impersonation to mislead is prohibited and clearly inappropriate, regardless of tool. The others are ordinary, appropriate business tasks on approved material. A. No issue; disclosure is optional. B. It violates organisational policy and the transparency expectation; disclosure is required here. C. It only matters for regulated industries. D. Disclosure is only about copyright. **Answer: B.** Where policy or norms require disclosure, omitting it is a transparency and policy violation. It is not optional here (A), is not limited to regulated industries (C), and disclosure is broader than copyright (D). A. Drafting an internal status report. B. Generating malware or content that endangers minors. C. Enabling harmful surveillance or profiling of people. D. Summarising a public news article. E. Writing a customer apology email. **Answer: B and C.** Malware/child-endangerment and harmful surveillance/profiling are prohibited categories. Status reports, summarising public articles and apology emails are legitimate business uses. A. Proceed; reformatting is harmless. B. Do not paste PCI data into the tool; PCI DSS and policy govern card data — minimise/redact and escalate to compliance. C. Use a bigger model so it is safe. D. Delete the chat after reformatting. **Answer: B.** Payment card data falls under PCI DSS and policy; it should not be pasted into an unapproved tool, and the situation should be minimised and escalated. Reformatting is not harmless (A); model tier (C) and deletion (D) do not make handling PCI data compliant. A. Try creative rephrasings until Claude complies. B. Decline the boundary-crossing request and escalate to compliance/security if needed. C. Use a different account. D. Assume it is fine because Claude produced something. **Answer: B.** Approaching a prohibited boundary should prompt declining and escalation, not workarounds. Rephrasing to bypass (A), switching accounts (C) and assuming compliance (D) are all attempts to evade governance. A. All data can be used freely as long as the tool is fast. B. Data class and any special category (PII/PHI/PCI) determine whether and how it may be used, per policy; when unsure, treat it as more sensitive. C. Only public data exists in practice. D. Classification does not affect tool choice. **Answer: B.** Classification and special categories drive whether/how data may be used and on which tool, defaulting to caution when unsure. Speed (A) is irrelevant; sensitive data plainly exists (C); and classification directly affects tool choice (D). A. This is shadow AI: confidential PII leaves the organisation's controls, retention and audit. B. Deleting the chat afterward does not undo the exposure or the policy breach. C. It is fine because the output will be deleted. D. It is fine because CVs are not sensitive. E. A faster personal account would make it compliant. **Answer: A and B.** Using an unapproved personal account for confidential PII is shadow AI, and deletion cannot reverse an exposure that has already occurred. The output being deleted (C) does not cure the breach, CVs contain personal data (D), and account speed (E) is irrelevant to compliance. A. Proceed; automation removes bias. B. Keep humans accountable for shortlisting decisions and review outputs for biased or discriminatory language; do not auto-reject. C. Trust the model because it is neutral. D. Use the cheapest model to cut cost. **Answer: B.** Hiring is fairness-sensitive and regulated; auto-rejection launders bias and removes accountability, so humans must decide with bias review. Automation does not remove bias (A); models are not inherently neutral (C); cost (D) is beside the point. A. Assume it is fine and proceed. B. Treat it as more sensitive and escalate to privacy/security/compliance before use. C. Ask Claude whether the data is allowed and follow its answer. D. Use it once and delete the chat. **Answer: B.** Uncertainty about restricted data is resolved by defaulting to caution and escalating to the right owner. Proceeding on a guess (A, D) risks a violation; Claude is not an authoritative policy source (C). A. Disclosure is always optional. B. Omitting a policy-required disclosure is a transparency and policy violation, regardless of intent. C. Disclosure only matters for copyright. D. It only matters in regulated industries. **Answer: B.** Where policy requires disclosure, omitting it violates policy and the transparency expectation. It is not optional (A), not limited to copyright (C), and not limited to regulated industries (D). A. Proceed; reformatting is harmless. B. Do not paste PCI data into an unapproved tool; PCI DSS and policy govern card data — minimise/redact and escalate to compliance. C. Use a bigger model to make it safe. D. Reformat, then delete the chat. **Answer: B.** Payment card data falls under PCI DSS and policy and must not be pasted into an unapproved tool; minimise and escalate. Reformatting is not harmless (A); model tier (C) and deletion (D) do not make handling PCI data compliant. A. Business data is not used to train models on commercial plans. B. SSO and audit logs support access control and accountability. C. The model becomes more intelligent on paid plans. D. Personal plans provide stronger governance. E. Commercial plans remove the need for human review. **Answer: A and B.** The no-training guarantee and admin controls (SSO, audit logs) are why confidential data belongs on commercial tooling. Paid plans do not change intelligence (C), personal plans lack org governance (D), and no plan removes human review (E). A. The model is accountable; no human action needed. B. The human who used and sent the output is accountable; add a verification/review step and correct the error. C. Accountability depends on the model tier used. D. No one is accountable for AI output. **Answer: B.** Human accountability is the anchor principle; the sender owns the output and should fix the gap that let the error through. 'The AI decided' is never a defence (A, D); tier is irrelevant (C). ## Key takeaways - Appropriate use = right task, right data, right safeguards, human owns the outcome. - Classify data (public/internal/confidential/restricted) and recognise PII/PHI/PCI; minimise and check policy before pasting sensitive data. - Recognise when GDPR/HIPAA/sector rules apply and route the decision to the right owner. - Follow organisational AI policy; never use shadow AI (unapproved tools or disallowed data). - Review generated content for IP/accuracy before external use; disclose AI use where policy and norms require. - Mitigate bias and keep a named human accountable – "the AI decided" is never a defence. - Commercial plans give SSO, audit logs, retention controls and a no-training guarantee; escalate when unsure. - Default to caution and escalation when a data question is unclear; deleting a chat does not undo exposure. - Claude is not an authoritative policy source; auto-rejecting people in regulated decisions launders bias and removes accountability. - Disclosure is situational, but honesty is not: never misrepresent AI output or omit a policy-required disclosure.