Terms and definitions

Representation of an entity, phenomenon, process or data, employing various AI algorithms to interpret, predict, or generate responses based on input. 

Machine learning model: mathematical construct that generates an inference or prediction based on input data or information.

Accountable: answerable for actions, decisions and performance within a well-defined scope of responsibility and potentially sanctionable.

Accountability: state of being accountable.

A set of instructions that guide a computer in performing specific tasks or solving problems.

Machine learning algorithm: algorithm to determine parameters of a machine learning model from data according to given criteria.

The process of measuring, evaluating and communicating something about a system or process, documentation, a product or an organisation. 

The overall aim of assurance is to verify and attest or assert that the system is operating as intended.

AI assurance: measures, evaluates and communicates the trustworthiness of AI systems.

The process of assessing against specific criteria with or without executing the artifacts, including model/system evaluation, capability evaluation, benchmarking, testing, verification, validation, as well as broader risk assessment and impact assessment.

AI model evaluation: the process of assessing an AI model against predefined specific criteria or general benchmarks (beyond accuracy). 

AI system evaluation: the process of assessing the functional accuracy/correctness and quality of an AI system against predefined specific criteria or general benchmarks (beyond model and correctness/accuracy).

AI capability evaluation: a comprehensive assessment of an AI model or system's overall capabilities, including both planned capabilities and unplanned, emerging, or dangerous capabilities. Unlike specific task-focused evaluations seen in AI model and system evaluation, capability evaluation seeks to understand the full range of an AI's capabilities. This includes evaluating how an AI might adapt or evolve beyond its initial training, identifying both beneficial emergent behaviours and potential risks that could arise from its autonomous operation or interaction with complex environments.

Treatment, behaviour or outcomes that respect established facts, beliefs and norms and are not determined or affected by favouritism or unjust discrimination (ISO TR24368).

Unfairness: unjustified differential treatment that preferentially benefits certain groups more than others.

A procedure that enables organisations to put their information classification scheme into practice by attaching classification labels to relevant information assets.

AI content labelling: applying visible content warnings to alert stakeholders to the presence of AI-generated content and its provenance.

Adapted from ISO 27002:2022 

and

Wittenberg

A qualitative or quantitative measure used to assess, compare, and track the performance or quality of a system, process, product or service.

Internal metrics measure the AI model or system itself (such as model complexity, explainability, training compute resources).

External metrics measure the behaviour and quality of the AI model or system (such as accuracy, response time, scalability).

Risk metrics measure the negative outcomes of using the AI system in a specific context (such as impacts on bias, privacy, security and compliance).

Impact metrics measure the broader effects of AI systems on users, groups and society.

An exercise, reflecting real-world conditions, that is conducted as a simulated adversarial attempt to provide a comprehensive assessment of the security capability of the AI system and organisation.

Adversarial testing: same.

Risk

Composite measure of an event’s probability of occurring and the magnitude of the consequences of the corresponding event. 

Note: The consequences of AI systems can be positive, negative, or both and can result in opportunities or threats.

Effect of uncertainty on objectives.

Chance of harm.

Organisation: Property of an organisation that appropriate activities and decisions are communicated to relevant stakeholders in a comprehensive, accessible and understandable manner.

System: Property of a system that appropriate information about the system is made available to relevant stakeholders. 

Appropriate information for system transparency can include aspects such as features, performance, limitations, components, procedures, measures, design goals, design choices and assumptions, (training) data sources and testing and evaluation methodologies, benchmarks, test cases and criteria.

Information embedded into digital content, either perceptibly or imperceptibly by humans, that can serve a variety of purposes, such as establishing digital content provenance or informing stakeholders that the contents are AI-generated or significantly modified.

AI-generated content watermarking: a procedure by which watermarks are embedded into AI-generated content. This embedding can occur at 2 distinct stages: during generation by altering a GenAI model's inference procedure or post-generation, as the content is distributed along the data and information distribution chain.