Quick Reference: Domain 6 — Governance, Risk, and Responsible Use

A task that suits Claude is not automatically permitted. Domain 4 asks whether a task suits Claude (shape, volume, checkability); this objective asks whether it should be handed over at all. A task can be structured, voluminous and checkable, and still be work that only a qualified human should do.

Four categories make a use inappropriate:

The substitution test: would you accept the work from a capable colleague with no professional qualification in the area, no access to the people's history, and no accountability for the outcome? Three grounds: qualification, history, accountability.

Avoid too much caution: most work is appropriate, and the exam marks too much caution as wrong as well. When a proposal mixes appropriate and inappropriate work, the right answer keeps the appropriate half. Clues: the internal-use excuse is never the right answer, and a quick sign-off at the end is not meaningful review. Hiring, housing and medical advice are examples of High-risk Areas. In these areas, the Usage Policy requires meaningful review by a qualified person with the authority to change the result. It also requires that the person affected is told AI was used. The usual safe answer is that Claude prepares the material and the person decides. Review never makes a prohibited use acceptable.

Recognising sensitive material has two parts. The obvious kind: names, addresses, dates of birth, account numbers, health information, anything about children. The half that people often miss is material that is sensitive because of who owns it or what was promised about it. Examples are a duty of confidence, commercially confidential figures and third-party material under contract. A pricing schedule with no personal data in it can still be the most restricted document you handle.

The four safeguards, in order. The right answer is the first safeguard in the list that works:

Why this order: you can do steps 1 and 2 yourself, today. Steps 3 and 4 wait for a different workspace or another person, so when removal is enough they cost time and add no protection. Abandoning the task is not one of the four safeguards and is almost never the right answer; the guide's own sample rationale says anonymisation enables the analysis.

Measures that seem to protect the data, but do not, fail for the same reason. Instructing Claude not to retain the data, declaring the upload confidential and deleting the conversation afterwards all act after the disclosure. The policy control is about what data leaves the organisation. That happened at upload. The safeguard happens before the upload or not at all.

Which product are you on? Consumer and commercial products answer differently whether they use your data for model training, and the details change. Questions test the professional behaviour: checking the current documentation for your product and your agreement. They never test a memorised retention period or default.

The workspace rule has real consequences. It is not a matter of manners. Three things apply to an organisation-provided workspace and to no personal account. The first is the organisation's agreement. Under it, the organisation decides what happens to its data, and Anthropic handles the data on the organisation's instructions. Commercial products do not train on inputs or outputs by default. The second is administration: the organisation's administrators control the workspace settings. The third is data handling: on Enterprise, the organisation sets custom retention timelines for its data. A paid personal plan with carefully chosen settings is still a personal account: what matters is whose agreement governs the data.

The shadow-AI response has two halves, and the wrong options reliably offer only the first. Do not copy the workaround, however pressing the deadline. Then report it to whoever owns the policy or the data. Report it because the disclosure that already happened needs to be known, and an unreported workaround continues. Having the colleague delete the chats is tidying after the event, not a fix for the disclosure.

A policy that says nothing about a use has not given permission for it. A policy that does not mention something has not answered the question; the right answer is asking the policy owner. When policy conflicts with convenience, the policy takes priority; when the policy is genuinely wrong or too slow, the right step is asking its owner to change it, not bypassing it.

A stricter policy, or no policy. When the organisation's policy is stricter than the workspace agreement, follow the policy. Continue the work within it, for example with the identifiers removed. When there is no AI policy at all, write a short rule: what may go in, what may not, and whom to ask. Take it to the people who run the organisation, and start when they agree it.

The examinable habits: know which account and plan you actually work in, know who administers it, follow the upload rules, know how to report a problem.

The accountability rule: the person whose name is on the work owns it completely. Saying Claude drafted it describes how the work was produced, like saying an assistant typed it. It moves no responsibility to the tool or to whoever configured the shared Project. The Usage Policy's high-risk requirements say the same: for advice or decisions in its High-risk Areas, a qualified person remains responsible for the accuracy and appropriateness of what is delivered or decided.

Disclosure depends on proportion, with two cases where it is a fixed requirement. Consumer-facing chatbots and external-facing AI agents must disclose that users are interacting with AI. A person who gets advice or a decision based on a High-risk AI Recommendation must be told that AI was used. Everything else follows the reliance test: what is the reader relying on? If it is the information alone, disclosure adds little; if it is the fact that you did the thinking (assessed work, a professional opinion), disclosure is owed. A condolence letter is relied on as your own feeling. A disclosure note cannot restore that, so you write it yourself. Both extremes are wrong options: nobody declares AI assistance on a routine internal email, and a rule to declare AI use everywhere hides the cases that matter among many unimportant ones. What a disclosure says: what AI was used for, how you reviewed it, and who is accountable for the result.

Fairness is a separate review from accuracy. An output can pass every fact-check and still characterise a group unfairly; the review needs someone who knows the people, reading specifically for how they are described.

Colleagues are stakeholders. When Claude takes over part of a process, the people doing that work deserve to know what changed and why; automating a shared task without telling the people who do it is an honesty question, not a quality one.

How to tell the domains apart: if an option's substance is checking, cross-referencing or validating, it answers Domain 2's question. A right answer for 6.4 is about ownership, disclosure, fairness or colleagues.

Current state, checked 30 September 2026. These are the facts behind this domain that are likely to change. They change, and none of them is worth memorising as an exam answer; what the exam gives marks for is the checking behaviour. Which product am I on, what does its documentation say today, and what does my organisation's agreement add?

• The Usage Policy (formerly the Acceptable Use Policy) is updated as Anthropic's technology and the associated risks evolve. By 11 October 2026 it had been reorganised around High-risk AI Recommendations in named High-risk Areas. It carries the chatbot disclosure requirement, the high-risk disclosure requirement and the high-risk accountability line that this domain uses (both disclosure requirements checked 29 October 2026).
• Training defaults differ by product. Two currently published articles share the title "Is my data used for model training?": the consumer one says consumer products train subject to the user's own model-improvement setting; the commercial one says commercial products do not train by default.
• Retention periods (from "How long do you store my data?"): deleted conversations leave back-end systems within 30 days; feedback submissions are kept for 5 years; opted-in training data is retained de-identified for up to 5 years; inputs and outputs flagged for Usage Policy violations for up to 2 years, with trust and safety classification scores for up to 7 years.
• The privacy policy is effective 10 September 2026, with Anthropic Ireland Limited as the controller for the EEA, UK and Switzerland (checked 30 September 2026).