CCAR-P - Governance, Safety & Risk Management (14% of the exam) - Section 5.5

Address ethical AI considerations (bias, fairness, transparency).

Identifying where an AI system could treat groups unfairly, testing for that bias, and being transparent with users about AI involvement and its limits. Candidates should choose measurable steps over statements of intent.

bias testingfairnesstransparencydisclosure of AI use

Practice question for this objective

Free sampleGovernance, Safety & Risk Managementmedium

A state benefits agency runs a Claude-based assistant that answers eligibility questions through its website chat and through an SMS text service. A survey of 600 users finds that 38 per cent believed they had been talking to a caseworker, and several treated an answer as a formal eligibility decision. Agency policy now requires every resident, on every channel, to know that they are dealing with an automated system, that its answers are not decisions, and how to reach a person. Which change best meets the requirement?

AState in the published terms of use and the privacy notice that eligibility questions are answered by an automated assistant.
BInstruct the model to confirm that it is an automated assistant whenever a resident asks whether they are speaking to a person.
COpen each conversation on both channels with a short notice of AI use, its limits and the route to a caseworker, then re-survey users. Correct
DShow a persistent banner above the website chat window stating that the assistant is automated and its answers are only guidance.

Disclosure of AI use must reach users proactively on every channel, state the system's limits, and be checked by measuring whether users understood it. Transparency is judged by what users actually understand, not by whether a disclosure exists somewhere. An opening notice is delivered in-band on every channel before the resident relies on any answer, and it can carry the limit (guidance, not a decision) and the escalation route together. Re-running the survey turns the change into a measurable step, showing whether the 38 per cent misunderstanding rate has dropped.

Why A is wrong: Formal notices are a familiar place for disclosures, so this feels compliant. It is wrong because residents rarely read them before asking a question, so the measured misunderstanding would persist, and it says nothing about answers not being decisions or how to reach a caseworker.

Why B is wrong: Honest answers to a direct question are necessary, which makes this appealing. It is wrong because the disclosure is reactive: the 38 per cent who assumed a caseworker never thought to ask, and the policy requires every resident to know, not only those who enquire.

Why C is correct: A notice at the start of every conversation reaches residents on the web and by SMS alike, and covers all three required elements: automation, that answers are not decisions, and how to reach a person. Repeating the survey measures whether the misunderstanding has actually fallen.

Why D is wrong: A visible banner is a strong disclosure on the web, so this is a reasonable proposal. It is wrong because the SMS service has no banner, so residents on that channel are left uninformed, and the banner omits the route to a person that the policy requires.

See more CCAR-P practice questions, answers explained.

Exam traps in Governance, Safety & Risk Management

Answers that look right on this material and are not. Each one is a distractor from a different question in the CCAR-P bank for this domain.

A standard notice on every score screen stating that results are AI-generated, may contain errors and should be reviewed by a person.
Why it is wrong: A visible AI notice is good practice, which makes this tempting. It is wrong because a generic disclaimer discloses no specific weakness, so customers still cannot tell candidates about the measured career-gap and language effects the vendor already knows about.
Set a separate flag threshold for each group so that interpreter-assisted patients and other patients are flagged at the same overall rate.
Why it is wrong: This is tempting because it closes the visible gap the board member raised and matches a well-known fairness definition, equal selection rates. It is wrong because the labelled cases show that true eligibility differs between the groups, so equal flag rates must either over-flag one group or keep missing eligible patients in the other, and it never measures the quantity the commitment names, which is how often eligible patients are missed.
Instruct the model in its system prompt to assign priorities consistently across districts, regardless of report length or style.
Why it is wrong: This targets the suspected cause cheaply, so it is tempting. It is wrong because an instruction is a statement of intent: it produces no monthly evidence of equitable service, and the council would not know whether the disparity had changed.

Examworthy is not affiliated with or endorsed by Anthropic. Original, blueprint-aligned practice material only.