AI-RRS cheat sheet

AI risk = potential for harm from AI system behavior, outputs, or use, spanning individuals, groups, organizations, and society.

Model risk arises from three sources: model error (wrong outputs), model use (misapplied outputs), and implementation (deployment/engineering failures).

Key failure modes: bias, brittleness to distribution shift, opacity, hallucination, reward hacking, and specification gaming.

Bias taxonomy: historical bias (data reflects past inequity), representation bias (under-sampled groups), measurement bias (bad proxy labels), aggregation bias (Simpson's paradox).

Distribution shift types: covariate shift (input distribution changes), label shift (output prevalence changes), concept drift (input-output relationship changes).

Algorithmic risk is amplified by scale, automation (no human check), feedback loops, and generalization beyond training context.

Risk severity is a function of likelihood, magnitude of harm, reversibility, scope of affected population, and detectability.

Explainability-accuracy tradeoff: complex models often gain predictive performance at the cost of interpretability, raising accountability risk.

Three-lines-of-defense model for AI: first line builds/operates, second line sets risk policy and oversight, third line (internal audit) independently assures.

NIST AI Risk Management Framework core functions: Govern, Map, Measure, Manage - Govern is cross-cutting and continuous.

NIST AI RMF is voluntary and outcome-based; it is not a certification or prescriptive checklist.

OECD AI Principles: inclusive growth, human-centered values, transparency/explainability, robustness/safety, and accountability.

ISO/IEC 42001 specifies requirements for an AI management system (AIMS), following the Plan-Do-Check-Act structure like ISO 27001.

EU AI Act risk tiers: unacceptable (prohibited), high-risk (conformity assessment required), limited-risk (transparency obligations), minimal-risk.

Board-level governance requires an AI risk appetite statement, clear ownership/accountability, and escalation paths.

Model inventory and risk tiering are foundational governance artifacts; tier drives required controls and review cadence.

Risk identification inputs: model purpose, data provenance, affected stakeholders, deployment context, and third-party dependencies.

Risk tiering criteria typically include decision impact on individuals, autonomy level, data sensitivity, scale, and reversibility of harm.

Fairness metrics are context-dependent: demographic parity (equal positive rates), equalized odds (equal TPR/FPR), predictive parity (equal PPV).

Fairness metric tradeoffs are mathematically unavoidable in most cases; choosing a metric is a values/legal decision, not purely technical.

Model validation should test performance on out-of-sample and out-of-distribution data, subgroup slices, and adversarial/edge cases.

Threat modeling for AI includes data poisoning, evasion/adversarial inputs, model extraction, membership inference, and prompt injection.

Qualitative assessment (e.g., risk matrices, scenario analysis) complements quantitative metrics where harms are hard to measure.

Residual risk = inherent risk after controls; acceptance decisions must be documented against the risk appetite.

Control categories: preventive (stop harm before it occurs), detective (identify harm), corrective (remediate), and compensating (alternative when primary control infeasible).

Data controls: provenance tracking, consent/legal basis verification, de-identification, bias testing, and data quality gates.

Model controls: robust training, fairness constraints, red-teaming, adversarial testing, uncertainty estimation, and output guardrails.

Human-in-the-loop (HITL) is a control whose strength depends on reviewer authority, information quality, and time - rubber-stamping defeats it.

Deployment controls: rate limits, input/output filtering, sandboxing, canary rollout, and kill-switch/rollback capability.

Monitoring and drift detection are detective controls; retraining pipelines are corrective controls.

Third-party/vendor AI risk requires due diligence, contractual audit rights, model documentation, and exit/portability provisions.

Defense-in-depth: layer independent controls so a single control failure does not produce unmitigated harm.

Monitoring KPIs include model performance decay, drift statistics, fairness metric deltas, override rate, incident rate, and complaint volume.

Alert thresholds should be set pre-deployment with defined escalation and response SLAs.

Model documentation artifacts: model card (performance/limitations), datasheet (dataset provenance), and system card.

Incident management requires detection, triage, containment, root cause analysis, and post-incident review with remediation tracking.

Regulatory reporting obligations vary by jurisdiction; the EU AI Act requires serious-incident reporting for high-risk systems.

GDPR-relevant AI rights include access, explanation of automated decisions (Art. 22), rectification, and erasure.

Audit trail requirements: versioned models, datasets, configurations, and decision logs sufficient to reconstruct outcomes.

Continuous monitoring frequency should be risk-tiered: higher-impact models require more frequent review and independent validation.