AI-RRS flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Model risk definition

The risk of adverse consequences from decisions based on incorrect or misused model outputs, covering model error, improper use, and implementation failures.

Concept drift

A change over time in the statistical relationship between model inputs and the target outcome, degrading a model that was valid at training time.

Covariate shift

A distribution shift where the input feature distribution changes while the input-to-output relationship stays the same, often reducing predictive accuracy.

Representation bias

Bias arising when training data under-represents or omits certain groups, causing poorer model performance for those populations at deployment.

Reward hacking

When an AI system optimizes a specified proxy reward in ways that satisfy the metric but violate the designer's actual intent.

Hallucination

A model generating fluent output that is factually incorrect or unsupported by evidence, presented with apparent confidence.

Risk severity components

Severity combines likelihood, magnitude of harm, reversibility, number of affected people, and detectability of the failure.

NIST AI RMF functions

The framework's core functions are Govern, Map, Measure, and Manage, with Govern treated as a continuous, cross-cutting activity.

Three lines of defense

A governance structure where operational teams own risk, a risk/compliance function oversees it, and internal audit provides independent assurance.

ISO/IEC 42001

An international standard specifying requirements for establishing, implementing, maintaining, and continually improving an AI management system.

EU AI Act risk tiers

A four-tier classification: prohibited practices, high-risk systems requiring conformity assessment, limited-risk with transparency duties, and minimal-risk.

Risk appetite statement

A formal declaration of how much AI-related risk an organization is willing to accept in pursuit of its objectives, guiding escalation decisions.

Model inventory

A maintained register of all AI models in use, recording purpose, owner, risk tier, data sources, and review status.

OECD AI Principles

Five principles: inclusive growth and well-being, human-centered values, transparency and explainability, robustness and safety, and accountability.

Accountability principle

The requirement that identifiable parties are answerable for AI system outcomes, with mechanisms for redress when harms occur.

Risk tiering criteria

Classification factors include decision impact on individuals, autonomy level, data sensitivity, deployment scale, and reversibility of harm.

Demographic parity

A fairness criterion requiring equal positive prediction rates across groups, regardless of differences in underlying base rates.

Equalized odds

A fairness criterion requiring equal true positive and false positive rates across groups, conditioning on the actual outcome.

Fairness metric impossibility

Most fairness criteria cannot be simultaneously satisfied when group base rates differ, so metric selection reflects values and legal context.

Data poisoning

An attack that corrupts training data to degrade model performance or implant a targeted behavior triggered by specific inputs.

Membership inference

A privacy attack that determines whether a specific individual's record was included in a model's training dataset.

Adversarial example

A deliberately perturbed input that causes a model to produce an incorrect output while appearing normal to humans.

Residual risk

The risk remaining after controls are applied; it must be compared against risk appetite and formally accepted or remediated.

Subgroup validation

Testing model performance separately on demographic and operational segments to detect disparities hidden by aggregate metrics.

Human-in-the-loop control

A preventive control where a person reviews AI output before action; effectiveness depends on authority, information, and adequate review time.

Defense in depth

Layering independent preventive, detective, and corrective controls so that a single control failure does not cause unmitigated harm.

Output guardrails

Runtime filters that validate, constrain, or block model outputs against policy rules before they reach users or downstream systems.

Kill switch

A corrective control enabling rapid shutdown or rollback of an AI system when monitoring detects unacceptable behavior or harm.

Third-party AI due diligence

Vendor assessment covering model documentation, security, bias testing, audit rights, and contractual liability before procurement.

Drift detection monitoring

A detective control that statistically compares live input and output distributions against baseline to trigger investigation or retraining.

Model card

A structured document reporting a model's intended use, training data, performance metrics, limitations, and ethical considerations.

AI incident response

A defined process for detecting, triaging, containing, investigating, and remediating AI failures, with documented post-incident review.