Fit vs Appropriateness
Two different questions on different evidence. Fit asks whether Claude would do the work well: the right shape, the right volume, output someone can check. Appropriateness asks whether the task should be handed over at all, and turns on stakes, authority and accountability.
Exam context: The recurring wrong answer reasons from quality: Claude would do the task quickly and well, so it must be fine. A task can be a perfect practical fit and still belong to nobody but a qualified human.
See also: 6.1 Identifying Appropriate and Inappropriate Use Cases
The Four Inappropriate Categories
Work requiring professional authority; consequential decisions about identifiable people; fabrication presented as real; and anything the Usage Policy or the organisation's own policy prohibits. The categories are narrow and specific, which is the point.
Exam context: Items reward naming which category makes a use inappropriate, not just that it feels risky. Everything outside the four categories is appropriate, and treating adjacent work as forbidden is the over-caution failure.
See also: 6.1 Identifying Appropriate and Inappropriate Use Cases
Substitution Test
The one-question appropriateness check: would you accept this work from a capable, well-meaning colleague with no professional qualification in the area, no access to the history of the people involved, and no accountability for the outcome? The AI Fluency framework calls the underlying skill Delegation.
Exam context: The three grounds are qualification, history and accountability. Speed, cost and output quality are not grounds; they are fit considerations wearing the test's clothes as distractors.
See also: 6.1 Identifying Appropriate and Inappropriate Use Cases
Usage Policy
Anthropic's canonical rules for what Claude may be used for, formerly called the Acceptable Use Policy. It requires consumer-facing chatbots to disclose that users are interacting with AI. For High-risk AI Recommendations, such as legal or medical advice to an individual, it requires review by a qualified person, who stays responsible for what is delivered or decided.
Exam context: The policy anchors two objectives: its prohibitions feed the fourth inappropriate category, and its accountability line is where the ownership anchor comes from. Expect older material to use the old name.
See also: 6.1 Identifying Appropriate and Inappropriate Use Cases
The Internal-Use Excuse
The argument that a use is acceptable because the results never leave the organisation. It appears across the domain, attached to inappropriate delegations and unprepared uploads alike, and it is never keyed.
Exam context: Keeping an output in-house restores neither accountability for the decision it embeds nor the affected person's route to challenge it, and where results go does not change what was disclosed at upload.
See also: 6.1 Identifying Appropriate and Inappropriate Use Cases
Sensitive Material
Anything that should not leave the organisation's hands unprepared. Two halves: the obvious inventory (names, addresses, dates of birth, account numbers, health information, anything about children) and material sensitive because of who owns it or what was promised about it.
Exam context: The under-recognised half is where working professionals fail: a document with no names in it can still be the most restricted document on the desk.
See also: 6.2 Applying Data Sensitivity, Regulatory and Privacy Considerations
Commercial Confidence
Sensitivity that comes from ownership and promises rather than personal identifiers: information held under a duty of confidence, commercially confidential figures, and third-party material governed by a contract that says how it may be handled.
Exam context: The classic distractor clears a document for upload because it contains no personal data. Confidentiality obligations bind regardless of whether a single person is named.
See also: 6.2 Applying Data Sensitivity, Regulatory and Privacy Considerations
The Safeguard Ladder
The ordered sequence of moves that lets sensitive work proceed: remove what the question does not need, anonymise or aggregate what it needs only in outline, use an authorised route, and finally ask whoever owns the data or the policy. Abandonment sits below the ladder.
Exam context: The order is the tested skill, because each rung preserves more of the work than the one below it. The keyed answer is the lowest rung that works; escalating past it reads as diligence but costs time for no protective gain.
See also: 6.2 Applying Data Sensitivity, Regulatory and Privacy Considerations
The Placebo Family
Safeguards that operate after the disclosure and therefore control nothing: instructing Claude not to retain the material, declaring the upload confidential, adding a confidentiality line to the prompt, and deleting the conversation afterwards.
Exam context: The guide's own sample rationale states the principle: instructing the model not to retain data does not satisfy the policy control. No member of the family is ever keyed; the keyed answer changes what goes in.
See also: 6.2 Applying Data Sensitivity, Regulatory and Privacy Considerations
Anonymise and Aggregate
The ladder's second rung: replacing individuals with cohorts and exact figures with ranges, so an analysis that needs the shape of the data can run without its identities. The guide's sample question turns on exactly this move.
Exam context: The enabling safeguard between two wrong extremes: uploading as-is and abandoning the analysis. Check the prepared version by asking whether anyone could be identified by combining what remains.
See also: 6.2 Applying Data Sensitivity, Regulatory and Privacy Considerations
Sanctioned Workspace
The Claude workspace an organisation provides and governs. Three things attach to it and reach no personal account: the organisation's agreement over how its data is handled, administrator-held settings, and organisation-level control over how long its data is retained.
Exam context: The rule is substance, not etiquette. Distractors dress it as cost allocation, manager oversight or courtesy; the keyed reasoning names the agreement and controls that stop at the workspace boundary.
See also: 6.3 Following Organisational AI Policies and Governance Standards
Personal Account
An account governed by the individual's own consumer relationship with Anthropic: their settings decide whether chats feed model improvement, and consumer retention mechanics apply. None of it is administered, seen or governed by the employer.
Exam context: The tempting variant is the paid personal plan with carefully chosen settings. Still a personal account: what matters is whose agreement governs the data, not how conscientiously the individual configured their own.
See also: 6.3 Following Organisational AI Policies and Governance Standards
Shadow AI
Work material flowing through accounts the organisation does not govern, typically a colleague's personal Claude account used because it is quicker or because sanctioned access has not arrived yet.
Exam context: The professional response has two halves: do not copy the workaround, and raise it with whoever owns the policy or the data. Distractors offer the first half alone, or deletion dressed as remediation.
See also: 6.3 Following Organisational AI Policies and Governance Standards
Policy Silence
The gap where an AI policy simply does not mention the material or the use in question. Silence means the question has not been answered, not that the answer is yes.
Exam context: Treating silence as permission converts a gap in a document into an unauthorised disclosure. The keyed move is asking the policy owner, which is an action, not a refusal, and it improves the policy for everyone.
See also: 6.3 Following Organisational AI Policies and Governance Standards
Controller and Processor
The regulatory vocabulary behind the sanctioned workspace: for commercial products, the customer organisation is Controller of the data its users submit, deciding what happens to it, while Anthropic acts as Processor, handling it only on the organisation's instructions.
Exam context: Lead with the plain-English gloss, the organisation decides and Anthropic follows its instructions; the formal labels are background. This split is what a personal account lacks.
See also: 6.3 Following Organisational AI Policies and Governance Standards
Accountability Anchor
The principle that the person whose name is on the work owns it completely: the sender of the email, the publisher of the report, the signer of the letter. Saying Claude drafted it is a statement about production method and transfers nothing.
Exam context: The distractor version spreads responsibility onto the tool or whoever configured the shared Project. Blame-the-tool options are never keyed; the keyed answer owns the error and corrects it to those affected.
See also: 6.4 Understanding the Ethical Implications of AI Usage
AI Disclosure
Telling an audience that AI was involved in producing work. One case is categorical under the Usage Policy: consumer-facing chatbots and external-facing AI agents must disclose that users are interacting with AI. Everything else runs on proportion via the reliance test.
Exam context: Both extremes are distractors: declaring AI assistance on every routine output, and never declaring it at all. Disclosure is owed when the reader is relying on who did the thinking or the judging, and it is not a shield for errors.
See also: 6.4 Understanding the Ethical Implications of AI Usage
Fairness Review
A separate read of any work that describes people, done by someone who knows them, looking only at how they are characterised. It sits apart from fact-checking because an output can pass every accuracy check and still frame a group unfairly.
Exam context: Options that stop at verified figures have answered the accuracy question while the ethical one is still open. Asking Claude to review its own draft for bias is a verification-flavoured placebo, not the review.
See also: 6.4 Understanding the Ethical Implications of AI Usage
Colleague Stakeholders
The people whose work changes when Claude takes over part of a process. They deserve to know what changed and why; quietly automating a shared task, or generating a colleague's contribution without telling them, is an honesty question, not a quality one.
Exam context: The stakeholder candidates most often miss. Stakeholder lists on the exam include colleagues, readers and the people described; they never include Anthropic, the model, or competitors.
See also: 6.4 Understanding the Ethical Implications of AI Usage