AAISM flashcards

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

Three lines model for AI

First line builds and operates AI and owns its risk; second line sets policy, appetite, and independent challenge; third line audits. Separating build from assurance prevents self-review of model risk.

AI system inventory

A register of every AI system recording owner, purpose, data classes, model type, autonomy level, and affected population. It is the prerequisite for tiering, assessment, and regulatory scoping.

AI risk tiering

Classifying each AI system by potential harm, such as prohibited, high-impact, limited, or minimal. Tier determines assessment depth, approval authority, documentation, and review cadence.

Human oversight control

A defined mechanism letting a named role review or override AI output, with authority, evidence standard, time window, and logging of each intervention. Oversight must be real and recorded.

NIST AI RMF functions

Govern, Map, Measure, and Manage. Govern establishes accountability and culture and wraps the other functions; Map frames context and impacts; Measure quantifies; Manage prioritizes and responds.

ISO/IEC 42001

An international standard specifying requirements for an AI management system, covering policy, roles, risk treatment, lifecycle controls, and continual improvement. It is certifiable by third-party auditors.

EU AI Act high-risk

AI used in employment, credit, education, essential services, or safety components faces obligations including risk management, data governance, logging, human oversight, and conformity assessment before market placement.

Model lifecycle gates

Mandatory approval checkpoints from intake through decommissioning: classification, data review, development testing, independent validation, deployment approval, monitoring, and retirement. Each gate needs documented sign-off.

Third-party AI due diligence

Vendor and foundation-model review requiring model documentation, training-data provenance, evaluation results, incident-notification terms, audit rights, and a viable exit or fallback plan.

AI governance council

A cross-functional body with defined decision rights that sets AI policy, reviews high-tier use cases, and approves risk acceptance above threshold. Acceptances are time-boxed and re-approved.

AI accountability owner

A named individual accountable for a specific AI system's performance, controls, incidents, and compliance. Diffuse ownership such as a team name is insufficient for governance.

Supervised learning

Training a model on labelled input-output pairs to minimize a loss function, enabling prediction on new inputs. Accuracy depends on label quality and representativeness of the training sample.

Unsupervised learning

Finding structure in unlabelled data through clustering, density estimation, or dimensionality reduction. Because there are no ground-truth labels, validation relies on stability and business interpretation.

Reinforcement learning

Training a policy that maximizes cumulative reward through interaction with an environment. Reward misspecification and unsafe exploration are the main control concerns.

Overfitting

A model fits training data noise so training error is low while validation error is high. Controls include held-out data, cross-validation, regularization, and leakage-free splits.

Data poisoning

Deliberate corruption of training data to degrade or manipulate model behavior. Controls include provenance verification, write-access restriction, dataset hashing, and label-distribution monitoring.

Prompt injection

Malicious input that causes a model to follow attacker instructions instead of the operator's. Mitigate by treating output as untrusted and enforcing least-privilege tool access.

Retrieval-augmented generation

A pattern where the model answers using documents retrieved at query time. Controls include permission-filtered retrieval, source citation, index freshness, and logging of retrieved chunks.

Model extraction attack

An adversary queries a model repeatedly to reconstruct an approximation or steal its behavior. Defenses include rate limiting, query monitoring, output perturbation, and contractual API restrictions.

Model inversion

Reconstructing sensitive training data or attributes from model outputs or confidence scores. Mitigate by limiting output detail, adding noise, and restricting access to predictions.

Hallucination rate

The proportion of outputs that are factually unsupported or fabricated, measured against a reference. It is a core evaluation metric for generative and RAG-based systems.

Model registry

A controlled repository of model artifacts, versions, metadata, and approvals. Immutable entries and signed deployments provide traceability from production behavior back to training inputs.

Data drift

A change in the distribution of model inputs compared with training data. It can degrade performance without any change to the underlying relationship between inputs and outcomes.

Concept drift

A change in the relationship between inputs and the target outcome over time, so previously learned patterns no longer hold. It requires retraining or rule revision.

Inherent versus residual risk

Inherent risk is likelihood times impact before controls; residual risk is what remains after control effectiveness. Risk is acceptable only when residual falls within stated appetite.

AI impact assessment

A documented analysis of how an AI system affects individuals and groups, covering affected populations, potential harms, mitigation, human review, and redress. Required for high-impact uses.

Algorithmic bias testing

Comparing error rates, selection rates, and outcomes across protected groups, plus examining proxy variables and label bias. Choosing a fairness metric is a policy decision, not purely technical.

AI incident response

Detect, contain by disabling or rolling back the model, assess harm and notification duties, remediate, and feed findings into risk assessment. It extends standard incident response with model-specific steps.

Model concentration risk

Dependence on a single model, vendor, or data supplier, where failure or change disrupts many processes. It belongs in the enterprise risk register with fallback and exit plans.

Risk appetite for AI

The amount and type of AI risk an organization is willing to retain, expressed as thresholds per tier. It anchors approval authority, escalation, and acceptance decisions.

AI red teaming

Adversarial testing that attempts to elicit harmful, biased, or unsafe model behavior before or after deployment. Findings must be tracked to remediation with named owners.

Privacy re-identification risk

The chance that anonymized or aggregated training data can be linked back to individuals through model outputs or auxiliary data. Assess before training on personal data.

Control effectiveness testing

Evidence-based verification that a stated AI control operates as designed, through sampling, re-performance, or metrics. Untested controls cannot justify a residual risk reduction.