Categories: Governance and Responsible AI

How to use this module: every scored answer should open by naming its category, defining it in one sentence, then enumerating what it involves — because the enumeration is what the note-taker writes down, and each named item is a potential rubric match. Category first, components second, your story third. These lessons are the category cards.

AI Governance

Definition: the organizational layer that keeps AI systems accountable — who owns each system, what policies apply, how risk is assessed, and how compliance is demonstrated.

What's involved:

1. An AI inventory — knowing every place AI is used, no shadow systems
2. Ownership and accountability — a named owner per system, answerable for it
3. Risk assessment per use case — keyed to whether it affects the public
4. Policy compliance — state GenAI guidance (EO N-12-23 / CDT), NIST AI RMF framing
5. Model transparency — every output explainable to an auditor
6. Review cadence and audit trail — decisions and changes are recorded and revisitable
7. The kill switch — capability can be disabled by decision, instantly

The opener: "This is fundamentally an AI-governance question — governance is the accountability layer: inventory, ownership, risk assessment, policy compliance, transparency, and audit. Let me take those in order for this case…"

Cite it when: the question mentions policy, oversight, "how do we control," compliance, "should we allow," or anything about the organization's relationship to AI rather than one system's internals.

Responsible AI

Definition: the design practice of building individual AI systems that are safe, fair, and trustworthy — governance's principles turned into engineering decisions. (Distinguish the pair: governance is the org layer; responsible AI is what you build into each system.)

What's involved:

1. Advisory vs. determinative — the system informs; humans (or auditable rules) decide
2. Human-in-the-loop — review paths for low-confidence and high-stakes outputs
3. Grounding and citations — answers traceable to authoritative sources
4. Output wording as architecture — "appears to meet," never "you are qualified"
5. Fairness — comparable people treated comparably; watch training-data bias
6. Measured accuracy before and after launch — no eval, no deploy
7. Fail-closed behavior — uncertainty routes to people, never to guessing
8. Microsoft's six principles by name when Azure-flavored: fairness, reliability & safety, privacy & security, inclusiveness, transparency, accountability

The opener: "I treat this as a responsible-AI design problem — advisory output, human review, grounded and cited answers, measured accuracy, fail-closed defaults. Here's how each applies…"

Cite it when: the question asks how to keep AI safe, accurate, fair, or appropriate for the public — anything where the concern is the behavior of one system toward people.

The pairing move

These two categories chain naturally, and chaining them is a senior move: "There's a governance layer — this system goes in our AI inventory with a named owner and a risk assessment — and a responsible-AI layer in the design itself: advisory output, human review, citations, measured accuracy." Two categories, ten scorable terms, fifteen seconds.