AnthropicClaude Certified Architect – Professional2026-105.5 Address ethical AI considerations (bias, fairness, transparency)
Explain AI assistance where people rely on it
Completed by the knowledge check
Learning objective
Provide transparency: disclose AI involvement, explain outputs and limitations, and cite sources.
Core idea
Transparency must help the person who relies on an answer understand what happened and what remains uncertain. A generic AI badge does not explain an individual recommendation; a source link does not disclose AI involvement; a polished rationale does not establish that the stated reason caused the result.
NIST AI RMF 1.0 separates transparency, explainability, and interpretability. Use that distinction to design a transparency specification: the notice people receive, the outcome explanation, the limitations presented at reliance points, and the sources available for verification. Each serves a different user need.
How it works
Disclose the actual role of AI
State what AI does in the workflow: retrieves information, drafts an explanation, recommends a disposition, or assists a qualified decision-maker. Name who finalizes the result and how a person can request review. Avoid a notice saying “AI makes no decisions” when its ranking controls which cases staff actually see.
Anthropic’s Usage Policy effective September 15, 2025 requires disclosure when model outputs are presented directly to individuals or consumers in its specified High-Risk Use Cases, at least at the beginning of each session. It separately requires consumer-facing chatbots, including external-facing or interactive AI agents, to disclose that users are interacting with AI rather than a human, at least at the beginning of each chat session. A human reviewer does not erase the AI contribution.
Keep the notice under application control so it appears before the person relies on the interaction. A generated sentence that sometimes omits the disclosure cannot reliably implement a session-start requirement. The policy version effective November 12, 2026 changes the disclosure wording and permits chatbot disclosure in the product interface; do not apply that future-effective alternative to the September 15, 2025 requirement.
The concrete artifact is a notice matrix: audience, entry point, AI role, final decision owner, required disclosure text, and test proving its presence. These are deployment choices implementing the policy, not Messages API fields that automatically disclose AI involvement.
Explain the outcome using checkable reasons
NIST’s transparency discussion includes when decisions were made and by whom; its interpretability discussion emphasizes the meaning of an output in its intended context. Translate that into an outcome explanation naming the applicable criterion, the case fact supporting it, the responsible decision-maker, and any unresolved information.
For an admissions recommendation, “the application does not contain the required prerequisite transcript” is checkable. “The model considered every factor carefully” is not. Distinguish missing information from a negative fact and distinguish a model recommendation from the human’s final determination. If a reviewer changes the recommendation, the explanation must describe the final decision rather than repeat the superseded draft.
Do not claim that a model-generated explanation proves its internal causal process. Verify the stated facts against the supplied records and approved rules. The user needs an intelligible basis for correction, not an unverifiable claim that fluent prose reveals everything the system did.
Present limitations at the reliance point
Document scope, input coverage, source dates, known weak populations or document formats, and what the service cannot establish from its inputs. NIST calls for transparency tailored to the role and knowledge of the audience. A technical system description can support auditors while a concise outcome notice helps an applicant correct a missing record.
Make limitations actionable. If the service lacks the institution’s assessment of overseas credentials, say the equivalence remains unresolved and identify the review route. A warning that “AI may make mistakes” does not tell someone which fact to challenge or whether the result is final. Avoid unsupported numerical confidence in place of a documented limitation.
Publish a system summary with intended use, decision criteria, evaluation scope, known limitations, and an accountable contact. Keep the version used for each released workflow. This is an organizational transparency artifact implementing the NIST approach; it does not certify fairness or replace evaluation.
Cite the material that supports the claim
For document-based explanations, the Claude API citations feature uses citations.enabled: true on supplied document blocks. The documentation requires citations to be enabled on all or none of the documents in a request. Citation support is text-based; it does not provide image citations. Do not promise a user a citation to visual evidence merely because the input is a PDF.
Returned text blocks can contain a citations list with cited_text, document_index, and source-location fields. Plain text uses character ranges, PDFs use page ranges, and custom content uses content-block ranges. Preserve the mapping to the source revision the application supplied; a citation to a subsequently replaced admissions rule needs its original context to remain intelligible.
Render the supporting passage and permitted source access with the explanation. A citation identifies where supporting text came from; reviewers still need to check that it supports the claim and that the source is applicable. A correct quotation from an obsolete rule is not a correct decision.
The citations documentation states that native citations and output_config.format structured outputs are incompatible in the same request and return HTTP 400 when combined. Choose a cited prose explanation when source inspection is the priority. If the workflow also requires a machine-readable decision record, separate that processing and validate the linkage; do not present a generated JSON source identifier as an equivalent native citation.
Key distinctions and trade-offs
Disclosure versus explanation: Disclosure identifies AI involvement. Explanation connects an outcome to criteria and case facts. Both can be required for the same interaction.
Provenance versus correctness: Source locations make checking possible. They do not establish source freshness, authorization, relevance, or the correctness of the final interpretation.
Transparency versus unrestricted disclosure: Tailor access to the audience. An applicant needs their own relevant records and reasons, not another applicant’s personal information or unrestricted internal logs. The system summary and user explanation can differ in detail while remaining consistent.
Architecture / decision scenario
An education provider handles 4,000 professional-course applications per month. Its portal uses Claude to draft prerequisite assessments and a qualified admissions officer finalizes each outcome. The service target gives staff two working days to answer challenges, and the institution must explain missing prerequisite findings without exposing other applicants’ records.
The product team proposes a footer saying “automated tools may be used” and one structured JSON response containing both the decision and native citations. The architect rejects the vague notice and incompatible request design. The portal states at session start that AI assists the assessment, identifies the admissions officer’s role, and provides the review route.
The explanation call supplies approved text documents with citations.enabled: true and returns cited prose. The application associates it with a separately validated decision record and source revisions. Staff verify the criteria and passages before release. If overseas credential equivalence remains unresolved, the explanation names that limitation and directs the applicant to specialist review.
The admissions owner signs off on notice coverage and explanation completeness; the platform owner verifies citation rendering and access controls. This costs additional processing and staff attention compared with one opaque score, but makes each disputed outcome checkable within the response target. The case-handling responsibility belongs to 5.5.4.
Common exam traps
- Treating a human-reviewed result as a reason to omit AI disclosure.
- Hiding the AI role in a vague footer instead of disclosing the interaction.
- Presenting a generated rationale as proof of internal model reasoning.
- Equating a citation with a verified, applicable source.
- Promising native citations for visual evidence in a PDF.
- Combining native citations and
output_config.formatin one request.
What to remember
Disclose the AI role before reliance, explain the final outcome through criteria and checkable facts, and make limitations actionable. Native citations support source inspection within their documented boundaries. Transparency succeeds when people can understand what happened and identify what needs correction.
Knowledge check
An insurer’s claims portal, operating under Anthropic’s Usage Policy effective September 15, 2025, which lists insurance claims among its High-Risk Use Cases, shows claimants coverage explanations that Claude drafts and a licensed adjuster approves before release. The product team proposes showing no AI notice, because a human approves every explanation. What should the portal do?
A benefits portal requires a prose explanation with native Claude API citations over supplied rule documents and a schema-enforced machine-readable decision record. The draft design is one request that sets citations.enabled: true on every document and also sets output_config.format with a JSON schema. Which design works?
output_config.format are incompatible: enabling citations on any document while setting a format returns HTTP 400, so a retry fails the same way and narrowing which documents cite does not help. A source identifier that the model writes into JSON is not an equivalent of a native citation. Two calls, one cited and one structured, with the link between them validated, give both outputs.An employment screening service drafts outcome letters with Claude, and a recruiter finalizes each one. Claude’s draft said a required certificate was missing. The recruiter found the certificate attached, but in a foreign format whose equivalence the service cannot assess, and changed the outcome to pending specialist review. What should the applicant’s explanation say?
A reviewer checks a cited explanation before a rejected permit application is released. The cited_text matches the supplied rule document word for word, but that document is a rule revision that was replaced last month, before this application was submitted. The permit authority confirms that the current revision governs this application. What should the reviewer do?
A lender’s evidence pack contains text-based PDF policy documents and scanned PDFs of applicants’ payslips that have no extractable text. The product team wants every explanation to cite the payslip page that supports it. What should the architect tell them about native citations?
page_location rangesReferences
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0)
- Anthropic, Usage Policy, effective September 15, 2025
- Anthropic, Usage Policy, effective November 12, 2026
- Anthropic, Citations