AI Governance Vocabulary

ACA Section 1557

The Affordable Care Act's nondiscrimination provision: bars discrimination in federally funded health programs and activities, and is applied to require covered entities to identify and address biased impacts of AI tools. [II.B.2]

Acceptable use

Defined boundaries for how the deployed system may and may not be used. [IV.C.1]

Acceptable use policy

Rules governing how staff may use third-party AI tools, including data they may input. [I.B.3; I.C.3]

Accountability

Identifiable people and processes are answerable for the system's outcomes. [I.A.4]

Accountable owner

The single named person answerable for a given AI system or decision. [I.B.1]

Administrative fine

A monetary penalty scaled to the severity of the breach and the firm's turnover. [II.C.5]

Adverse impact ratio

A fairness metric comparing a protected group's selection/advancement rate to a reference group's; a ratio below the accepted threshold (e.g. the four-fifths rule) signals potential disparate impact. [III.B.3]

Agentic architecture

Systems where AI plans and takes actions with tools, increasing autonomy and risk. [IV.A.3]

AI Basic Act (South Korea)

The second comprehensive national AI law (effective Jan 2026) — regulates 'high-impact AI' and generative-AI transparency. [II.C.1]

AI governance committee

A cross-functional body that sets policy, reviews high-risk use cases and oversees the program. [I.B.1]

AI impact assessment (AIA)

A structured evaluation of an AI system's potential impacts and the mitigations for them. [III.A.2]

AI incident

An event where an AI system causes or risks harm, failure or non-compliance. [III.C.4]

AI Liability Directive (AILD)

A proposed EU measure on fault-based (negligence) liability for AI harm — withdrawn by the Commission in 2025; distinct from the revised Product Liability Directive's no-fault, defect-based regime. [II.B.4]

AI life cycle

The end-to-end stages from use-case assessment through retirement of an AI system. [I.C.1]

AI literacy

A baseline understanding of AI capabilities, limits and risks — increasingly a legal expectation. [I.B.3]

AI management system (AIMS)

A structured set of policies and processes to govern AI, certifiable under 42001. [II.D.3]

AI maturity model

A staged view of governance capability, from ad hoc to optimised. [I.B.4]

AI Office

The EU body overseeing GPAI and coordinating enforcement. [II.C.5]

AI RMF Playbook

Companion guidance suggesting concrete actions for each function and category. [II.D.2]

AI washing

Overstating or fabricating AI capabilities in marketing. [II.B.3]

Allocative harm

Harm from an AI system withholding or granting opportunities or resources unfairly (a loan, a job, a place). [I.A.2]

Artificial Intelligence

A machine-based system that, for explicit or implicit objectives, infers from its inputs how to generate outputs such as predictions, content, recommendations or decisions. [I.A.1]

Audit

A structured, often independent review of a system against defined criteria. [III.C.3]

Automated decision-making (ADM)

A decision based solely on automated processing, restricted where it produces legal or similarly significant effects. [I.C.2; II.A.3]

Automation bias

Over-reliance on automated outputs — a risk that human oversight must actively counter. [II.C.3]

Autonomy

The degree to which a system acts and decides without real-time human intervention. [I.A.3]

Awareness program

Ongoing communication that keeps AI risks, policies and norms salient across the organisation. [I.B.3]

Benchmarking

Comparing performance or risk against references or baselines. [III.A.4]

Bias / interpretability testing

Assessing fairness across groups and the explainability of outputs. [III.B.3]

Biometric data

Data derived from physical or behavioural traits used to identify a person (face, fingerprint, voice). [II.A.4]

Brittleness

Failure on inputs slightly different from the training conditions. [III.C.5]

Business context

The organisational objectives, constraints and stakeholders surrounding the system. [III.A.1]

Catastrophic forgetting

Losing earlier capabilities when a model is retrained on new data. [III.B.4]

Classic (predictive) AI

Models that predict or classify, often more interpretable and task-specific. [IV.A.2]

Cloud / on-premise / edge

Where the model runs, trading scalability, control, latency and privacy. [IV.A.3]

Code of conduct

A voluntary commitment to responsible-AI practices that minimal-risk providers are encouraged to adopt. [II.C.1]

Conformity assessment

The process of demonstrating a high-risk system meets legal requirements before it is placed on the market. [II.C.2]

Conformity requirements

The legal pre-market checks a high-risk system must satisfy. [III.C.1]

Contextual risk

Risk that depends on how and where a system is actually used. [IV.B.1]

Continuous monitoring

Ongoing observation of an AI system's performance and behaviour in production. [III.C.2]

Contractual safeguards

Clauses on data use, transparency, audit, liability and incident notice that allocate AI risk. [I.C.3]

Copyright

Exclusive rights in original creative works — relevant to both training inputs and AI outputs. [II.B.1]

Counterfactual explanation

An explanation showing what minimal change would alter the outcome — e.g., 'approved if income were higher.' [III.C.6]

Crisis communication

Prepared messaging for incidents and failures. [IV.C.6]

Cross-functional collaboration

Structured joint work across legal, technical, business, ethics and risk functions on AI decisions. [I.B.2]

Dark pattern

An interface designed to manipulate users into choices against their own interest. [II.B.3]

Data availability

Whether suitable data exists to run or adapt the system. [IV.A.1]

Data catalog

A managed inventory of datasets with metadata, sources and lineage. [III.B.2]

Data controller

The entity that determines the purposes and means of processing personal data. [II.A.3]

Data dependency

An AI system's behaviour is determined by training data, so data flaws become system flaws. [I.A.3]

Data governance

Policies and processes for sourcing, quality, use, access and retention of data. [III.B.1]

Data governance policy

Rules for how data is sourced, classified, used, retained and protected — extended here to AI training and use. [I.C.2]

Data leakage

Test information contaminating training, which inflates performance. [III.B.4]

Data localization

A legal requirement that certain data be stored and processed only within the country where it was collected. [II.A.3]

Data minimization

Processing only personal data that is adequate, relevant and limited to what is necessary for the purpose. [II.A.2]

Data quality

The accuracy, completeness, consistency and timeliness of data. [III.B.1]

Data residency

Where data is physically stored — often a business or policy choice rather than a legal mandate. [II.A.3]

Data sovereignty

The principle that data is subject to the laws of the country in which it is located. [II.A.3]

Data-use clause

Terms on whether and how the vendor uses your data, including for training. [IV.B.2]

Datasheet (for datasets)

Documentation of a dataset's source, composition and limitations. [III.A.5]

Deactivation (kill switch)

The ability to stop or disable an AI system promptly. [IV.C.7]

Deceptive practice

A material representation or omission likely to mislead a reasonable consumer. [II.B.3]

Decision rights

Explicit rules about who can approve, escalate or veto an AI decision. [I.B.2]

Decommissioning

The controlled retirement of an AI system, including data handling. [IV.C.7]

Demographic parity

A fairness metric requiring outcomes to be distributed similarly across groups. [III.B.3]

Deployer

The party that uses an AI system under its own authority in a real context. [I.B.5; II.C.6]

Deployer impact assessment

An assessment of risks in the specific deployment context. [IV.B.1]

Deployment audit

A periodic review of a deployed system against criteria in your environment. [IV.C.3]

Deployment governance

Applying policies and controls to the operational use of an AI system. [IV.C.1]

Design defect

A foreseeable risk that a safer, reasonable alternative design would have avoided. [II.B.4]

Design log / decision record

A record of key design choices and the rationale behind them. [III.A.5]

Developer / provider

The party that builds an AI system or places it on the market under its name. [I.B.5]

Digital Markets Act (DMA)

EU competition law imposing obligations on large 'gatekeeper' platforms (e.g. no self-preferencing, interoperability, data-combination limits) — distinct from the DSA's content/transparency focus. [II.B.3]

Digital Services Act (DSA)

EU law on online platforms requiring recommender-system transparency, profiling/ad-targeting disclosure, and action against illegal content and systemic risks such as disinformation. [II.B.3]

Disparate impact

A neutral practice that disproportionately harms a protected group, potentially unlawful absent justification. [I.A.4; II.B.2]

Disparate treatment

Intentional differential treatment based on a protected characteristic. [II.B.2]

Distribution shift

When live data diverges from training data, degrading performance. [III.B.4]

Distributor

A supply-chain party (other than provider or importer) that makes a system available, checking marking and conformity downstream. [I.B.5]

Diversity of perspective

Including varied roles, backgrounds and affected viewpoints to reduce blind spots. [I.B.2]

Downstream harm

Harm arising further along from the system's outputs or decisions. [IV.C.5]

Downstream provider

A party that builds an AI system on top of a GPAI model. [II.C.4]

DPIA

A data protection impact assessment, required for processing likely to result in high risk to individuals. [II.A.3; III.A.2]

Drift

Degradation as live inputs or relationships diverge from training. [IV.C.2]

Equalized odds

A fairness metric requiring comparable error rates (false positives/negatives) across groups. [III.B.3]

Escalation path

A defined route for raising a contested or higher-risk AI decision to a more senior or specialised body. [I.B.2]

Ethics

The overarching commitment to build and use AI in ways that respect human rights and societal values — the umbrella over the other responsible-AI principles. [I.A.4]

Ethics by design

Building ethical requirements into the system from the outset, not bolting them on later. [I.C.1; III.A.3]

EU AI Act Article 4

Requires providers and deployers to ensure a sufficient level of AI literacy among staff and others operating AI on their behalf. [I.B.3]

Evaluation (TEVV)

The broader assessment of a system's overall performance and quality against defined metrics. [III.B.3]

Evaluation metrics

The quantitative measures used to judge performance and fairness. [III.B.5]

Expert system

A rule- and knowledge-based system that emulates the decision-making of a human specialist — a classic non-ML form of AI. [I.A.1]

Explainability

The reasons for a specific output can be understood by an appropriate audience. [I.A.4]

Explicit consent

A clear, specific, affirmative agreement, frequently required to process special-category data. [II.A.4]

External communication plan

A plan for informing external stakeholders about the AI system and any incidents. [IV.C.6]

Failure to warn

Inadequate instructions or warnings about a product's risks and limitations. [II.B.4]

Fairness

Avoiding unjust bias and ensuring comparable treatment across individuals and groups. [I.A.4]

Fallback process

The manual or alternative process used when the AI is disabled. [IV.C.7]

Feature

An individual measurable property or characteristic of the data used as input to a model (e.g. a house's square footage or location). [I.A.1]

Feature flag (feature toggle)

A switch that turns specific functionality or model behaviour on or off dynamically at runtime without redeploying the system — enabling controlled experiments, gradual rollout and immediate shut-off. [IV.C.7]

Fine-tuning

Further training a model on specific data to fit a task. [IV.A.3]

Fit-for-purpose

Data suitable in content and quality for the model's intended use. [III.B.1]

Foundation model

A large model trained on broad data that can be adapted (e.g., fine-tuned) to many downstream tasks. [I.A.1]

Four-fifths (80%) rule

A US disparate-impact screen: a selection rate below 80% of the highest group's signals adverse impact. [III.B.3]

Framework Convention on AI

The Council of Europe's 2024 treaty — the first legally binding international agreement on AI, open to non-European states. [II.C.1]

FRIA

A fundamental-rights impact assessment required of certain high-risk deployers. [III.A.2; IV.B.1]

Function creep

Gradual expansion of a system's use beyond the original scope. [IV.C.5]

Fundamental-rights impact assessment (FRIA)

A deployer assessment required for certain high-risk uses. [II.C.6]

Gap assessment

A structured review comparing current policies against AI-specific risks. [I.C.2]

General-purpose AI (GPAI) model

A model with broad capability usable across many downstream tasks. [II.C.4]

Generative AI

Models that produce new content — text, image, audio, code — by learning the statistical structure of large training datasets. [I.A.1; IV.A.2]

Global vs local explanation

Global explains the model's overall behaviour; local explains one specific output. [III.C.6]

Go/no-go decision

A deliberate release decision to proceed or halt, based on whether gate criteria are met. [I.C.1]

Govern / Map / Measure / Manage

The four core functions of the AI RMF. [II.D.2]

High-impact AI

South Korea's term for AI affecting life, safety or fundamental rights — the analogue of the EU's high-risk tier. [II.C.1]

High-risk AI

Systems posing significant risk to health, safety or rights, subject to strict obligations. [II.C.1]

Hub-and-spoke governance

A central AI governance committee setting policy and reviewing high-risk cases, with embedded contacts ('spokes') handling triage in each business unit. [I.B.2]

HUDERAF / HUDERIA

The Council of Europe's Human Rights, Democracy and the Rule of Law Assurance Framework (and its HUDERIA impact-assessment methodology) — proportional impact assessments plus stakeholder engagement to operationalise the Framework Convention's principles. [II.C.1]

Human authorship requirement

The rule in several regimes that copyright requires a human author, limiting protection for AI-generated works. [II.B.1]

Human oversight

A person's ability to monitor, intervene in or override an AI system's operation, scaled to its risk and autonomy. [I.A.3; II.C.3]

Human-centricity

AI is designed to respect human autonomy, rights and wellbeing. [I.A.4]

Hyperparameter

A configuration setting chosen before training that controls how the model learns (e.g. learning rate, number of layers); tuned, not learned. [I.A.1]

IEEE 7000

IEEE standard model process for addressing ethical concerns during system design — embeds individual/societal values into the life cycle with traceability, distinct from ISO 22989's terminology focus. [II.D.3]

Importer

A party that places an AI system from a non-EU provider on the EU market and must verify its conformity first. [I.B.5]

Importer / distributor

Parties that bring or make AI available on the market and verify its conformity. [II.C.6]

Incident log

A record of issues and incidents arising in deployment. [IV.C.4]

Incident management

The defined process for detecting, responding to and learning from AI failures. [I.C.1]

Indemnification

A contractual shift of liability — for example for IP infringement. [IV.B.2]

Inferred data

Attributes a model derives about a person that were never directly provided — often missed by pre-AI privacy policy. [I.C.2]

Inferred sensitive data

Special-category attributes a model derives from non-sensitive inputs. [II.A.4]

Instructions for use

Provider documentation enabling deployers to operate the system correctly and safely. [II.C.3; III.C.6]

Intake process

A defined route by which new AI use cases enter governance review. [I.B.2]

Integration risk

Risk arising from how the model connects to your systems and data. [IV.C.3]

Intellectual-property exposure

The risk that an AI system trains on or reproduces copyrighted or proprietary material, or leaks trade secrets through its outputs. [I.A.3]

IP in AI

Questions of rights in training data and ownership of AI-generated outputs. [I.C.2]

ISO 31000:2018

The general enterprise risk-management standard (principles + framework + process) for any organisation and any risk; ISO/IEC 23894 adapts it specifically for AI. [II.D.3]

ISO/IEC 22989

The standard defining AI concepts and terminology. [II.D.3]

ISO/IEC 42001

The certifiable AI management system (AIMS) standard. [II.D.3]

ISO/IEC 42005

Guidance on conducting AI system impact assessments. [II.D.3]

Issue / risk register

A maintained list of open issues and risks with owners. [IV.C.4]

Issue management

The process for handling problems that arise during use. [IV.C.1]

Joint controllers

Two or more parties that together determine the purposes and means of processing (GDPR Art. 26); they must agree an arrangement allocating their respective responsibilities. [II.A.3]

Label

The target or output value a supervised model learns to predict (e.g. the house's price). [I.A.1]

Lack of robustness

Sensitivity to noise, edge cases or adversarial inputs. [III.C.5]

Lawful basis

A legal ground required to process personal data (e.g., consent, contract, legitimate interests under GDPR). [II.A.1]

Legitimate interests

A lawful basis permitting processing necessary for an interest not overridden by individuals' rights and freedoms. [II.A.1]

Licence terms

Contractual conditions governing whether data or a model may be used for training. [II.B.1]

Licensing agreement

The contract governing use of a vendor's AI model or system. [IV.B.2]

LIME / SHAP

Post-hoc techniques that estimate how much each input feature contributed to a prediction. [III.C.6]

Limited-memory AI

An AI system that retains recent data to inform its decisions; nearly all useful AI today is limited-memory. [I.A.1]

Limited-risk AI

Systems subject mainly to transparency obligations such as disclosure. [II.C.1]

Lineage

The record of data's movement and transformations through the pipeline. [III.B.2]

Localization

Restricting a system's operation to certain regions or contexts. [IV.C.7]

Machine Learning (ML)

A subset of AI in which models learn patterns from data rather than executing explicitly programmed rules. [I.A.1]

Maintenance schedule

A planned cadence for updates, checks and retraining. [III.C.2; IV.C.2]

Make-vs-buy

The decision to build and own versus license a third-party model. [IV.B.3]

Manufacturing defect

A departure from the intended design — for AI, e.g., flawed data or training producing unintended behaviour. [II.B.4]

Market-surveillance authority

A national body enforcing the AI Act and policing the market. [II.C.5]

Membership inference

An attack that determines whether a specific person's data was in the training set. [III.C.3]

Minimal-risk AI

Systems with no specific obligations beyond generally applicable law. [II.C.1]

Misalignment

A gap between the objective the system optimises and the outcome the organisation actually intends. [I.A.2]

Misuse

Use contrary to the intended purpose or acceptable-use rules. [IV.C.5]

Model / data drift

Degradation as data or relationships change over time. [III.C.5]

Model card

A standard document describing a model's purpose, data, performance and limits — common release-gate evidence. [I.C.1; III.A.5; III.C.1]

Model drift

Degradation of a model's performance over time as data or the world diverge from what it was trained on. [I.A.3; III.C.2]

Model inversion

An attack that reconstructs sensitive training data from a model's outputs. [III.C.3]

Model memorisation

When a model reproduces training data, creating privacy and IP exposure. [I.C.2]

Model provenance

The documented origin and history of a model and its training data, used to assess third-party risk. [I.C.3]

Model selection

Choosing an architecture or model fit for the purpose, risk and explainability needs. [III.A.3]

Multimodal model

A model that handles multiple data types such as text, image and audio. [IV.A.2]

Narrow (weak) AI

AI built for a specific task or bounded set of tasks. All AI in real-world use today is narrow AI. [I.A.1]

Natural language processing (NLP)

Techniques that let systems understand, interpret and generate human language. [I.A.1]

NIST AI RMF

A voluntary US framework (2023) for managing AI risks and improving trustworthiness. [II.D.2]

OECD AI Principles

Intergovernmental, values-based principles for trustworthy AI adopted in 2019 and updated in 2024. [II.D.1]

OECD.AI Policy Observatory

The OECD's hub tracking national AI policies, metrics and trends. [II.D.1]

Opacity (black box)

The difficulty of inspecting or explaining how a model reached a particular output. [I.A.3]

Operational controls

Design-time mechanisms — guardrails, thresholds, oversight points — that govern behaviour. [III.A.3]

Overfitting

A model that learns training noise and fails to generalise to new data. [III.B.4]

Parameter

An internal value the model learns from data during training — the weights adjusted to fit the data; not set by hand. [I.A.1]

Performance requirements

The accuracy, latency and reliability the use case demands. [IV.A.1]

Phased applicability

Staggered dates on which different obligations take legal effect. [II.C.5]

Policy floor

A minimum baseline of controls that applies everywhere, even under decentralised or federated governance. [I.B.4]

Post-market monitoring plan

A provider's plan to track a system's real-world performance after release. [III.C.6; IV.C.4]

Post-mortem

A structured, blameless review of an incident's causes. [III.C.5]

Precision / recall / F1

Performance (accuracy) metrics — precision = share of positive predictions that are correct, recall = share of actual positives caught, F1 = their harmonic mean; they measure correctness, not fairness across groups. [III.B.3]

Privacy

Personal data used by an AI system is collected and processed lawfully and minimally, with individuals' rights protected. [I.A.4]

Privacy by default

The most privacy-protective settings apply automatically, without the user having to act. [II.A.2]

Privacy by design

Embedding data-protection measures into systems and processes from the outset. [II.A.2]

Privacy-enhancing technologies (PETs)

Techniques such as pseudonymisation, anonymisation, differential privacy and federated learning that reduce privacy risk. [II.A.2]

Probabilistic output

Outputs are statistical likelihoods, not guaranteed results — the same prompt can yield different answers. [I.A.3]

Probability/severity matrix

A tool ranking risks by their likelihood and their impact. [III.A.4]

Problem framing

Articulating the underlying need and whether AI is the appropriate solution. [III.A.1]

Processor

An entity that processes personal data on the controller's behalf (e.g., a cloud or model vendor). [II.A.3]

Production monitoring

Observing a deployed system's behaviour in real use. [IV.C.2]

Prohibited (unacceptable-risk) AI

Practices banned outright due to unacceptable risk to rights and safety. [II.C.1]

Prompt injection

Crafting inputs that override a generative model's instructions or guardrails. [III.C.3]

Proportionality

Calibrating the strength of controls to the level of risk and organisational capacity. [I.B.4; III.A.2]

Proprietary model

A model an organisation builds, owns and controls. [IV.B.3]

Proprietary vs open-source model

Vendor-controlled vs openly available models, differing on control, support and transparency. [IV.A.2]

Protected characteristic

A trait (race, sex, age, disability and others) that law shields from discrimination. [II.B.2]

Provenance

The origin and source history of data — who, where, and under what rights. [III.B.2]

Provider

Builds or places an AI system on the market under its name; bears the primary obligations. [II.C.6]

Provider notification

Informing the provider of incidents for their post-market duties. [IV.C.4]

Provider obligations

The heavier legal duties borne by those who build or place AI on the market. [IV.B.3]

Proxy variable

A feature correlated with a protected attribute that can reproduce its effect. [II.B.2]

Purpose limitation

Personal data collected for one purpose may not be freely reused for an incompatible new purpose. [II.A.1]

Quality management system (QMS)

Documented processes ensuring the ongoing compliance of high-risk AI. [II.C.3]

RACI

A matrix assigning who is Responsible, Accountable, Consulted and Informed for each activity. [I.B.1]

Reactive machine

An AI system that responds only to current input and keeps no memory of past interactions (e.g., IBM Deep Blue). [I.A.1]

Rebuttable presumption of defectiveness

A legal assumption that a product is defective unless the manufacturer proves otherwise (e.g. that it met applicable safety and legal requirements) — it shifts the burden, it is not a permanent or automatic finding. [II.B.4]

Record-keeping (logging)

Automatic logs enabling traceability of the system's operation. [II.C.2]

Red teaming

Adversarial testing that actively tries to make the system fail or misbehave. [III.C.3; IV.C.3]

Redress channel

A route for affected people to complain or seek correction. [IV.C.6]

Release readiness

A go/no-go evaluation that a system is fit and compliant for production. [III.C.1]

Representational harm

Harm from a system reinforcing stereotypes, demeaning or erasing groups, even without a direct allocation decision. [I.A.2]

Representativeness

Whether the data reflects the population the model will affect. [III.B.1]

Reproducibility

The ability to recreate the model and results from documented inputs. [III.B.2; III.B.5]

Requirements gathering

Eliciting functional, legal and ethical requirements before building. [III.A.3]

Residual risk

Risk remaining after mitigations, requiring formal sign-off. [IV.B.1]

Retraining trigger

A predefined condition that prompts a model refresh. [III.C.2]

Retrieval-augmented generation (RAG)

Grounding a model's outputs in retrieved, authoritative data. [IV.A.3]

Revised EU Product Liability Directive

An updated regime expressly covering software and AI and easing proof of defect. [II.B.4]

Risk tiering

Sorting use cases into risk bands so the weight of controls scales with potential harm. [I.B.4]

Risk tolerance

The amount and type of risk an organisation is willing to accept in pursuit of its objectives. [I.B.4]

Risk-management system

A continuous process to identify, evaluate and mitigate risks across the AI lifecycle. [II.C.2]

Risk-mitigation hierarchy

A preference order: eliminate > reduce > transfer > accept. [III.A.4]

Robustness testing

Checking behaviour under edge cases, noise and adversarial inputs. [III.B.3]

Role-based training

Training content tailored to what each audience must know and do. [I.B.3]

Rollback plan

A predefined way to revert or disable the system if needed. [III.C.1]

Root-cause analysis

Determining the underlying cause to prevent recurrence. [III.C.4]

Safety and reliability

An AI system performs as intended within safe limits and behaves consistently, including on unexpected inputs. [I.A.4]

Safety component

An AI system whose failure could endanger health or safety — a high-risk route independent of Annex III (e.g., autonomous vehicles, aviation, medical devices). [II.C.1]

Scalability risk

The amplification of a small error or bias into systemic harm once a model is deployed at scale. [I.A.2]

Secondary use

Using a system for a purpose beyond its intended one. [IV.C.5]

Security

AI systems and their data are protected from unauthorised access, manipulation and attacks such as poisoning or evasion. [I.A.4]

Security testing

Probing for vulnerabilities such as data poisoning, prompt injection and model theft. [III.C.3; IV.C.3]

Self-aware AI

A hypothetical AI possessing consciousness and awareness of its own existence — the highest, theoretical level. [I.A.1]

Serious-incident reporting

A legal duty to notify authorities of significant AI incidents. [III.C.4]

Service-level agreement (SLA)

Committed performance and availability terms. [IV.B.2]

Shadow AI

Unsanctioned use of AI tools by staff outside approved channels — a key risk that awareness programs aim to reduce. [I.B.3; I.B.4]

Soft law

Non-binding standards that shape norms and frequently harden into binding law over time. [II.D.1]

Special-category data

Personal data needing extra protection: health, race/ethnicity, religion, political opinions, sexual orientation, biometric and genetic data. [II.A.4]

Stage gate

A checkpoint a project must clear before moving to the next lifecycle stage. [I.C.1]

Stakeholder disclosure

Telling affected parties about AI use and its implications. [IV.C.6]

Stakeholder mapping

Identifying who is affected and how, to surface risks. [III.A.4]

Sub-processor

A vendor's own downstream supplier that processes data or provides an AI component on its behalf. [I.C.3]

Substantial modification

Changing a system enough that the deployer takes on provider-like obligations. [I.B.5]

Success metrics

Defined measures of whether the system achieves its intended purpose. [III.A.1]

Systemic risk

Risk from highly capable GPAI that could have large-scale impact, triggering extra obligations. [II.C.4]

Technical documentation

Records demonstrating a high-risk system meets requirements (design, data, testing). [II.C.2; III.A.5]

Test (TEVV)

Controlled checks with predefined inputs and expected outputs confirming specific functions behave correctly under known conditions — narrow functionality, not overall performance. [III.B.3]

Test data

A dataset held untouched until the end, used once for a final, unbiased simulation of real-world performance (a.k.a. evaluation data). [III.B.4]

Test report

A record of test plans, metrics, results and pass/fail decisions. [III.B.5]

TEVV

Test, evaluation, verification and validation processes for AI. [III.B.3]

Text and data mining (TDM) exception

A legal carve-out permitting some analysis of copyrighted works, frequently subject to a rights-holder opt-out. [II.B.1]

Theory-of-mind AI

A hypothetical, advanced AI that understands the beliefs, intentions and emotions of others. [I.A.1]

Third-party risk

Risk introduced by external vendors, models or services embedded in your operations. [I.C.3]

Threat modeling

Systematic identification of attack vectors and abuse cases. [III.C.3]

Three lines model

A governance accountability structure: the business owns and manages risk (1st line), risk/compliance sets policy and oversees (2nd line), and internal audit gives independent assurance to the board (3rd line). There is no fourth line. [I.B.1]

Tiered disclosure

Giving each audience the appropriate level of information. [III.C.6]

Total cost of ownership

The full cost of building, securing and maintaining a model. [IV.B.3]

Training data

The dataset the model learns from — it adjusts the model's internal parameters (weights). [III.B.4]

Training documentation

Records of the data, configuration and process used to train the model. [III.B.5]

Training-data summary

A required public summary of the content used to train a GPAI model. [II.C.4]

Transparency

Individuals must be clearly informed about how their personal data is processed and with what consequences. [I.A.4; II.A.1]

Transparency obligation

A duty to disclose specified information about an AI system. [III.C.6]

Triage

Classifying an incident by severity to drive the response. [III.C.4]

Trustworthy AI

AI that is lawful, ethical and robust, reflecting agreed responsible-AI principles. [II.D.1]

Trustworthy-AI characteristics

The properties the RMF promotes — valid/reliable, safe, secure, accountable/transparent, explainable, privacy-enhanced, fair. [II.D.2]

Unfair practice

A practice causing substantial, unavoidable consumer harm not outweighed by countervailing benefits. [II.B.3]

Use case

The specific problem, users and context an AI system is built to serve. [III.A.1]

Use-case context

The business, data, ethical and organisational setting of a deployment. [IV.A.1]

User / end-user

The individual interacting with the system or affected by its output. [I.B.5]

User training

Preparing staff to use and oversee the system correctly. [IV.C.1]

Validation (TEVV)

Confirming the right system was built — that it meets the real-world need and intended use. [III.B.3]

Validation data

A held-aside dataset never used to adjust internal parameters; used repeatedly during development to give an unbiased performance check that guides hyperparameter/design tuning and the decision to keep tuning or stop. [III.B.4]

Validation testing

Confirming the model meets its requirements on held-out data. [III.B.3]

Vendor due diligence

Structured assessment of a supplier's AI practices before and during engagement. [I.C.3]

Vendor-update management

Tracking and validating model changes pushed by a vendor. [IV.C.2]

Verification (TEVV)

Confirming the system was built right — that it meets its design specifications and requirements. [III.B.3]

Workforce readiness

Whether staff are trained and prepared to use and oversee the system. [IV.A.1]

Zero-shot learning (ZSL)

A model recognising or categorising classes it was never explicitly trained on. [I.A.1]