Every term, defined

AI System#

Engineered system that generates outputs such as content, forecasts, recommendations or decisions for a given set of human-defined objectives.

(ISO/IEC 22989:2022 3.1.4)

Artificial Intelligence (AI)#

A system that, for a given set of human-defined objectives, generates outputs such as predictions, recommendations, classifications, or decisions, and that exhibits the capacity to perform tasks commonly associated with human cognition (e.g., perception, reasoning, learning, decision-making) under varying conditions.

(Adapted from ISO/IEC 22989:2022 3.1.4; modified to add capacities associated with human cognition. ISO/IEC 22989 3.1.3 defines artificial intelligence as the discipline.)

Continuous Learning#

Incremental training of an AI system that takes place on an ongoing basis during the operation phase of the AI system life cycle.

(ISO/IEC 22989:2022 3.1.9)

Determination Boundary#

A defined and bounded collection of hardware and software items, together with its declared interfaces, on which the AI applicability determination of Section 1.2.1 is performed. Framework requirements attach to learned components within the boundary and to artifacts whose correctness criterion is fidelity to a learned component; obligations on the integrating system attach where the respective requirements specify. Adapted from the AI/ML constituent concept of the EASA Artificial Intelligence Concept Paper Issue 2; modified to a domain-agnostic determination unit for framework applicability.

Format Conversion#

A model transformation that translates a model between framework representations (e.g., PyTorch to ONNX to TensorRT) to enable execution on target runtimes. Subject to revalidation per AI-4.7.

Human-Machine Teaming#

Integration of human interaction with machine intelligence capabilities.

(ISO/IEC 22989:2022 3.3.3)

Hyperparameter#

Characteristic of a machine learning algorithm that affects its learning process and is selected prior to training.

(ISO/IEC 22989:2022 3.3.4)

Inference#

Reasoning by which conclusions are derived from known premises.

(ISO/IEC 22989:2022 3.1.17)

Knowledge Distillation#

A model transformation that trains a smaller “student” model to approximate the behavior of a larger “teacher” model, producing a compact model with comparable functional behavior. Subject to revalidation per AI-4.7.

Learned Artifact#

An artifact whose values are determined or improved from data by machine learning (ISO/IEC 22989:2022 3.3.5), whether before deployment, through retraining (3.3.10), or through continuous learning during operation (3.1.9), and that serves as the specification of system behavior without a documented derivation constituting a correctness argument for that behavior. A trained model (3.3.14) is the paradigm case. A model designed for continuous learning is a learned artifact at every point in the operation phase, including prior to its first operational update. Data-fitted parameters within explicitly written model structures whose derivation constitutes a correctness argument are not learned artifacts (Table 12b). Framework-defined to make the Section 1.2.1 determination decidable, drawing on ISO/IEC 22989 process terminology.

Learned Component#

A component whose behavior is specified by one or more learned artifacts. Framework requirements attach to learned components within the determination boundary and to artifacts whose correctness criterion is fidelity to a learned component (Section 1.2.1). Framework-defined.

Machine Learning#

Process of optimizing model parameters through computational techniques, such that the model’s behavior reflects the data or experience.

(ISO/IEC 22989:2022 3.3.5)

Machine Learning Model#

Mathematical construct that generates an inference or prediction based on input data.

(ISO/IEC 22989:2022 3.3.7)

Model#

Physical, mathematical or otherwise logical representation of a system, entity, phenomenon, process or data.

(ISO/IEC 22989:2022 3.1.23)

Parameter (Model Parameter)#

Internal variable of a model that affects how it computes its outputs (weights in a neural network).

(ISO/IEC 22989:2022 3.3.8)

Pruning#

A model transformation that removes model weights, neurons, or layers determined to contribute minimally to model output, reducing model size and computational cost. Subject to revalidation per AI-4.7.

Quantization#

A model transformation that reduces the numerical precision of model weights and/or activations (e.g., FP32 to INT8) to reduce memory footprint and computational cost, typically for execution on resource-constrained hardware. Subject to revalidation per AI-4.7.

Retraining#

Updating a trained model by training with different training data.

(ISO/IEC 22989:2022 3.3.10)

Trained Model#

Result of model training.

(ISO/IEC 22989:2022 3.3.14)

Training / Model Training#

Process to determine or to improve the parameters of a machine learning model, based on a machine learning algorithm, by using training data.

(ISO/IEC 22989:2022 3.3.15)

Mission-Critical AI#

AI outputs affect operational success but not human safety.

(Framework-defined, Section 1.2.2)

Operational Support AI#

AI supports operations but does not drive critical decisions.

(Framework-defined, Section 1.2.2)

Safety-Critical AI#

AI outputs directly affect human safety or system survivability.

(Framework-defined, Section 1.2.2)

ODD Declaration#

The formal documentation of an AI system’s Operational Design Domain, including environmental conditions, input characteristics, operational scenarios, and constraints under which the system is designed to function as validated. Per AI-1.0.

(Framework-defined; aligned with SAE J3016 ODD concept.)

ODD Evolvability#

The declared property of an Operational Design Domain that specifies whether and how the ODD may change over the system’s operational life through continuous learning, retraining, or reconfiguration. Systems with non-evolvable ODD declarations operate within a frozen specification; evolvable ODDs require additional controls per AI-13.

(Framework-defined.)

Operational Claim#

Single-sentence summary statement that opens an ODD declaration, summarizing the function, subject population, operational conditions, user role, and regulatory regime of the certified system.

(Framework-defined, AI-1.0)

Operational Design Domain (ODD)#

The operational conditions, inputs, users, subjects, regulatory context, and constraints under which an AI system is designed and validated to operate. The ODD defines the scope of the certification claim and the envelope within which reliable behavior is expected. See also ODD Declaration; ODD Evolvability; Operational Claim.

(Framework-defined, Section 1.2.3 and AI-1.0; concept aligned with ISO 34503 and SAE J3016)

Subject Population#

The population of persons or entities that an AI system’s decisions act upon, distinct from the user population that operates the system. Examples include patients in a medical decision support system, pedestrians and cyclists in a vehicle perception system, or obstacle classes in an autonomous platform. See also User Population.

(Framework-defined, AI-1.0)

User Population#

The set of human operators, controllers, supervisors, or other authorized personnel who interact with the AI system to direct, monitor, or oversee its operation. Distinct from Subject Population, which refers to entities the AI system acts upon.

(Framework-defined; see Subject Population.)

Ground Truth#

Value of the target variable for a particular item of labelled input data.

(ISO/IEC 22989:2022 3.2.7)

Input Data#

Data for which an AI system calculates a predicted output or inference.

(ISO/IEC 22989:2022 3.2.9)

Label#

Target variable assigned to a sample.

(ISO/IEC 22989:2022 3.2.10)

Production Data#

Data acquired during the operation phase of a deployed AI system for which a deployed system calculates a predicted output or inference.

(ISO/IEC 22989:2022 3.2.12)

Test Data#

Data used to assess the performance of a final model.

(ISO/IEC 22989:2022 3.2.14)

Training Data#

Data used to train a machine learning model.

(ISO/IEC 22989:2022 3.3.16)

Validation Data#

Data used to compare the performance of different candidate models.

(ISO/IEC 22989:2022 3.2.15)

Bias#

Systematic difference in treatment of certain objects, people, or groups in comparison to others.

(ISO/IEC 22989:2022 3.5.4)

Concept Drift#

Change in the relationship between input data and the target variable over time (i.e., the decision boundary moves), requiring relabeling of training data and model retraining. See also Continuous Validation.

(ISO/IEC 5338:2023 6.4.14 / ISO/IEC 22989:2022 5.11.9, tailored)

Data Drift#

Change in the statistical characteristics of production data over time compared to the training data distribution, which can degrade model prediction accuracy. See also Continuous Validation.

(ISO/IEC 5338:2023 6.4.14 / ISO/IEC 22989:2022 5.11.9, tailored)

Hallucination#

The production of confidently stated but erroneous or false content. (NIST AI 600-1 §2.2 (Confabulation))

Model Drift#

Degradation in model performance over time due to data drift, concept drift, or other factors affecting the relationship between inputs and outputs.

(Tailored)

Out-of-Distribution#

Input data that falls outside the statistical distribution of the training data, for which model predictions may be unreliable or undefined. (Tailored from ISO/IEC 23053:2022) See ISO/IEC 23053:2022 §8.7 (Operation), which discusses data drift, concept drift, and generalization affecting model performance.

Overfitting#

Creating a model which fits the training data too precisely and fails to generalize.

(ISO/IEC 23053:2022 3.2)

Performance Drift#

Degradation of AI model performance metrics over operational time, typically arising from data drift or concept drift. Continuous monitoring per AI-4.4 and AI-4.6 detects performance drift; periodic revalidation per AI-4.5 quantifies it. See also Continuous Validation.

(Framework-defined; complements Data Drift and Concept Drift.)

Continuous Validation#

Ongoing monitoring process to verify that AI models continue performing satisfactorily during operation, including detection of data drift, concept drift, and performance degradation. See also Data Drift; Concept Drift; Performance Drift.

(ISO/IEC 5338:2023 6.4.14, tailored)

Control Point#

A defined verification or authorization gate in the AI system lifecycle through which an artifact, decision, or change must pass before proceeding to the next stage. Used in cross-references between AI-4.3 (Model Maintenance Criteria), AI-4.7 (Deployment Format Validation), and AI-7.4 (Model Integrity).

(Framework-defined.)

Deployment-Format Model#

An AI model that has been transformed from its trained form for execution on operational hardware through processes such as quantization, pruning, knowledge distillation, or framework conversion. Subject to revalidation per AI-4.7.

(Framework-defined.)

Fairness Metrics#

Quantitative measures used to assess whether an AI system produces equitable outputs across different groups, operational contexts, and mission scenarios. Includes metrics for disparate impact, equal opportunity, demographic parity, and performance consistency across operational conditions.

(Tailored from ISO/IEC TR 24027)

System Safety Process#

The system-level safety analysis and assessment activity through which a project identifies hazards and produces the hazard analyses that software-level standards consume, conducted per applicable domain practice (e.g., SAE ARP4754/ARP4761 safety assessment in civil aviation; ISO 26262-3 hazard analysis and risk assessment in automotive; ISO 14971 risk management for medical devices; MIL-STD-882E system safety in defense; the NASA safety program hazard analysis process, from which NASA-STD-8739.8 derives software safety criticality). Distinct from the software standard this framework supplements (e.g., DO-178C, NPR 7150.2D, ISO 26262-6), which consumes system safety outputs to determine software criticality but does not itself conduct the system-level hazard analysis.

(Framework-defined, AI-1.9 and Section 1.2.5)

Test#

Execution of the system or component under controlled conditions with defined inputs, producing quantitative measurements evaluated against pass/fail criteria.

Validation#

Confirmation, through the provision of objective evidence, that the requirements for a specific intended use or application have been fulfilled.

(ISO/IEC 22989:2022 3.5.18)

Verification#

Confirmation, through the provision of objective evidence, that specified requirements have been fulfilled.

(ISO/IEC 22989:2022 3.5.17)

Autonomous Action#

System action taken without requiring human approval or intervention at the time of execution.

(Tailored)

Controllability#

Property of an AI system that allows a human or another external agent to intervene in the system’s functioning.

(ISO/IEC 22989:2022 3.5.6)

Fallback Behavior#

Predefined safe system response invoked when AI outputs are unreliable, unavailable, or flagged as out-of-distribution.

(Tailored)

Human Override#

Capability for human operators to supersede or reject AI recommendations or autonomous actions, with documented rationale.

(Tailored)

Reliability#

Property of consistent intended behavior and results.

(ISO/IEC 22989:2022 3.5.9)

Resilience#

Ability of a system to recover operational condition quickly following an incident.

(ISO/IEC 22989:2022 3.5.10)

Robustness#

Ability of a system to maintain its level of performance under any circumstances.

(ISO/IEC 22989:2022 3.5.12)

Adversarial Attack#

Deliberate manipulation of AI system inputs, training data, or model parameters intended to cause incorrect or harmful outputs.

(ISO/IEC 5338:2023 6.4.3)

Adversarial Example#

An input crafted by perturbing a legitimate input with imperceptible or low-magnitude modifications designed to cause AI model misclassification or incorrect output. Distinct from Adversarial Attack, which refers to the broader class of attack techniques. Per AI-7 and AI-12.6.

Adversarial Robustness#

Ability of an AI system to maintain correct and safe operation despite adversarial inputs or attacks.

(ISO/IEC 22989:2022 3.5.12, tailored)

Data Poisoning#

Injection of malicious data into training datasets to influence model behavior in attacker-controlled ways.

(ISO/IEC 5338:2023 6.4.3)

Input Manipulation#

Crafting of inputs designed to cause model misclassification or incorrect outputs while appearing valid to human observers.

(ISO/IEC 5338:2023 6.4.3)

Model Inversion#

Attack technique that reconstructs sensitive training data by analyzing model outputs.

(ISO/IEC 5338:2023 6.4.3)

Model Theft#

Extraction of model parameters or behavior through systematic querying to replicate proprietary AI capabilities.

(ISO/IEC 5338:2023 6.4.3)

Confidence#

Quantified measure of certainty or probability associated with an AI output or prediction.

(Tailored)

Explainability#

Property of an AI system to express important factors influencing the AI system results in a way that humans can understand.

(ISO/IEC 22989:2022 3.5.7)

Graceful Degradation#

System capability to maintain reduced functionality when AI is impaired.

(Tailored)

Predictability#

Property of an AI system that enables reliable assumptions by stakeholders about the output.

(ISO/IEC 22989:2022 3.5.8)

Transparency#

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

(ISO/IEC 22989:2022 3.5.15)

Trust Calibration#

Alignment between operator trust in AI and actual AI reliability.

(Human factors)

Anonymization#

The process of rendering personal data such that the data subject is no longer identifiable, by any means reasonably likely to be used. Regulatory standards for anonymization vary across jurisdictions, including GDPR Recital 26, HIPAA Safe Harbor at 45 CFR 164.514(b)(2), and HIPAA Expert Determination at 45 CFR 164.514(b)(1). See also De-identification.

(Framework-defined; aligned with GDPR Recital 26.)

Data Controller#

The natural or legal person, public authority, agency, or other body which, alone or jointly with others, determines the purposes and means of the processing of personal data. (GDPR Article 4(7))

Data Minimization#

The principle and practice of collecting, processing, and retaining only the personal data strictly required for the AI system’s operational task, with the minimum sufficient subset kept and the remainder discarded or replaced with derived features. Per AI-10 and Section 5.4.1. (Aligned with GDPR Article 5(1)(c) and ISO/IEC 27559.)

Data Processor#

A natural or legal person, public authority, agency, or other body which processes personal data on behalf of the controller. (GDPR Article 4(8))

Data Subject#

The identified or identifiable natural person to whom personal data relates. (GDPR Article 4(1))

De-identification#

The process of modifying personal data to reduce the likelihood of associating it with specific individuals, encompassing techniques such as masking, generalization, perturbation, and aggregation. Distinct from Anonymization in regulatory frameworks (HIPAA, CCPA) where de-identification implies a quantified risk threshold. Per AI-10 and Section 5.4.2. See also Anonymization; Re-identification Risk.

Differential Privacy#

A mathematical framework for quantifying the privacy risk of a computation over a dataset, characterized by a privacy budget parameter (epsilon) that bounds the contribution of any individual record to the output. (Framework-defined; see Dwork et al. 2006 foundational literature) Full citation: Dwork, C., McSherry, F., Nissim, K., & Smith, A. (2006). “Calibrating Noise to Sensitivity in Private Data Analysis,” in Theory of Cryptography Conference (TCC 2006), Lecture Notes in Computer Science, vol. 3876, Springer, pp. 265-284.

DPIA (Data Protection Impact Assessment)#

A structured assessment of the privacy risks of a data processing activity, required under GDPR Article 35 and analogous regulations for high-risk processing. For AI systems processing personal data, DPIA is typically performed at design time and updated when the data processing materially changes. See also Privacy Impact Assessment (PIA).

Lawful Basis#

The legal justification for processing personal data under an applicable privacy framework. Examples include the six bases enumerated in GDPR Article 6, HIPAA permitted uses and disclosures under 45 CFR 164.502, and CCPA business purposes.

(Aligned with GDPR Article 6.)

Membership Inference#

An attack class in which an adversary determines whether a specific data record was included in the training set of an AI model, typically by exploiting model confidence patterns or output behavior. Per AI-10 and Section 5.4.5.

Personal Data#

Any information relating to an identified or identifiable natural person. The precise scope varies by regulatory framework. Examples include GDPR Article 4(1), HIPAA Protected Health Information under 45 CFR 160.103, and CCPA “personal information” under Cal. Civ. Code 1798.140. (Aligned with GDPR Article 4(1).)

Privacy Impact Assessment (PIA)#

Systematic assessment of the privacy risks associated with a processing activity and the mitigation measures applied. Under GDPR Article 35 and for high-risk processing, referred to as a Data Protection Impact Assessment (DPIA).

(Framework-defined, AI-10.6; GDPR Article 35)

Privacy-Enhancing Technique (PET)#

Technical method designed to reduce privacy risk through data or computation transformation, including anonymization, pseudonymization, differential privacy, federated learning, secure multi-party computation, and homomorphic encryption. (Framework-defined, AI-10.4) See ISO/IEC 27559:2022, “Information security, cybersecurity and privacy protection Privacy enhancing data de-identification framework.”

Pseudonymization#

The processing of personal data in such a manner that the data can no longer be attributed to a specific data subject without the use of additional information, where that additional information is kept separately and subject to technical and organizational measures. (GDPR Article 4(5))

Re-identification Risk#

The likelihood that personal data which has been deidentified or anonymized can be associated back to specific individuals through techniques such as auxiliary data linkage, attribute inference, or longitudinal tracking. Per AI-10 and Section 5.4.2.

Special Categories of Personal Data#

Categories of personal data subject to additional protection under GDPR Article 9 (racial or ethnic origin, political opinions, religious or philosophical beliefs, trade union membership, genetic data, biometric data for identification, health data, sex life or sexual orientation) or analogous categories under other frameworks. Examples include HIPAA PHI and CCPA “sensitive personal information.”

(Aligned with GDPR Article 9.)