AI ethics names the outcomes we want: fair, transparent, accountable, human-centered systems. AI governance is the operating model that turns those outcomes into repeatable decisions — who decides, what gets checked, what evidence is kept, and what happens when a system fails. This guide bridges the two.
Ethics documents state principles: fairness, transparency, accountability, privacy, human oversight, safety. Governance documents state controls: risk tiers, review boards, model documentation, monitoring thresholds, incident response. Teams that publish principles without controls end up with unenforceable policy. Teams that publish controls without principles end up optimizing for compliance instead of outcomes.
| Ethical principle | Operational control | Evidence |
|---|---|---|
| Fairness | Disparate-impact testing across protected classes before release and on a recurring cadence. | Test plan, subgroup metrics, sign-off memo. |
| Transparency | Model card, data card, decision-reason surfacing to affected users. | Published model card version, adverse-action reason codes. |
| Accountability | Named model owner, named business owner, escalation path, override policy. | RACI, decision log, override sample review. |
| Privacy | Data minimization, purpose binding, retention schedule, DPIA for high-risk uses. | DPIA, data lineage, retention job logs. |
| Human oversight | Human-in-the-loop for high-impact decisions; auto-referral thresholds. | Referral rates, reviewer QA scores. |
| Safety & robustness | Red-team plan, drift monitoring, kill-switch procedure. | Red-team report, monitoring dashboard, tabletop exercise. |
Nearly every mature program runs on a 3-tier scheme. Tiering is what turns a general policy into proportional effort — a low-risk marketing model does not need the same controls as a credit decisioning system.
Controls attach to lifecycle stages, not to teams:
Fairness is not one metric. Pick the definition that matches the harm: demographic parity for equal opportunity contexts, equalized odds when both false positives and false negatives carry weight, calibration when the score itself is used across groups. Document the choice and the tradeoff — a system tuned for demographic parity will not also satisfy calibration except in degenerate cases.
Model cards are for reviewers. Users need something different: a plain-language reason for the decision that affects them, a path to contest it, and a person who can review the contest. In regulated contexts (credit, employment, housing, insurance) this is a legal requirement; in the rest it is what earns trust.
"The model decided" is not an accountability statement. For every deployed system there is a model owner (usually engineering) and a business owner (the function that benefits). Both are named, both sign the launch memo, and both are on the escalation path when the system misbehaves.
Human-in-the-loop only works when the human has (a) time, (b) information, and (c) authority to overturn. Otherwise it is rubber-stamping. Design the review surface so the reviewer sees the top drivers of the decision, the confidence level, and the historical override rate for similar cases. Track reviewer agreement rate — if it is 99%, review is theater.
Generative systems add three failure modes classical ML did not have: data leakage (the model regurgitates training or prompt content), hallucination (confident false claims), and prompt injection (untrusted input steers the model). Governance responses:
A mature program has a single inventory of AI systems, each tagged with a tier, an owner, a launch date, a monitoring status, and a link to its model card. It runs a governance council on a fixed cadence. It has a documented incident-response process that has been exercised at least once. And it can, on any given day, answer three questions in under a minute: what AI do we run, who owns it, and what is failing right now?