ISO/IEC 22989:2022 — AI Concepts and Terminology (standard)

ISO/IEC 22989:2022 — Information technology — Artificial intelligence — Artificial intelligence concepts and terminology

Tag: S-2022-07-iso-iec-22989 Type: report (international standard — primary normative text) Author(s): ISO/IEC JTC 1/SC 42, Artificial Intelligence Date of source: 2022-07 (first edition) Date ingested: 2026-08-17 Authority weight: high — the normative ISO/IEC vocabulary standard for AI; the terminology base referenced by ISO/IEC 42001 and the wider SC 42 corpus Raw file: S-2022-07-iso-iec-22989.pdf (single-user licence to Redstrata Ltd / Paul Miles, ISO Store order OP-1086858, downloaded 2026-08-16; copying and networking prohibited — wiki pages summarise, they do not reproduce)

What it claims

Establishes terminology for AI and describes concepts in the field, for use in developing other standards and supporting communication among stakeholders; applicable to all types of organizations. It has no normative references and defines 117 terms across seven groups: AI general (3.1, 35 terms), data (3.2, 15), machine learning (3.3, 17), neural networks (3.4, 10), trustworthiness (3.5, 18), natural language processing (3.6, 18) and computer vision (3.7, 4). Its central definition: an AI system is an “engineered system that generates outputs such as content, forecasts, recommendations or decisions for a given set of human-defined objectives” (3.1.4).

Clause 5 covers AI concepts: narrow vs general AI (weak/strong AI dismissed as philosophical); the agent paradigm; knowledge as technical content (not cognition) in the data–information–knowledge hierarchy; symbolic vs subsymbolic vs hybrid AI; the data processes pipeline (acquisition through labelling) and data roles (training, validation, test, production); ML approaches (supervised, unsupervised, semi-supervised, reinforcement, transfer, continuous learning, retraining, data/concept drift, catastrophic forgetting); example algorithms (neural networks incl. FFNN/RNN/LSTM/CNN, Bayesian networks, decision trees, SVM); a 7-level autonomy/heteronomy/automation scale (Table 1, levels 0–6) with a note that “autonomous” is a misnomer for current AI since systems are not self-governing; IoT and cyber-physical systems; trustworthiness characteristics (robustness, reliability, resilience, controllability, explainability, predictability, transparency, bias and fairness — distinguishing technical bias from unfairness); verification vs validation; jurisdictional issues; societal impact risk factors; and a stakeholder role taxonomy (AI provider, producer, customer, partner — including AI auditor and AI evaluator — subject, and relevant authorities).

Clause 6 gives a non-prescriptive AI system life cycle: inception; design and development; verification and validation; deployment; operation and monitoring; continuous validation (for continuous-learning systems); re-evaluation; retirement — each with example processes and repeated pointers to ISO/IEC 23894 for risk management. Clause 7 gives the functional view (model at the core; prediction → decision → action; AI outputs are error-prone and probabilistic; “AI systems do not understand”). Clause 8 describes the AI ecosystem in functional layers (AI systems, AI function, ML, engineering approaches incl. expert systems and logic programming, big data, cloud/edge computing with three cloud-edge training patterns, resource pools incl. ASICs). Clause 9 covers fields (computer vision, NLP and its component tasks, data mining, planning); clause 10 example applications (fraud detection, automated vehicles, predictive maintenance). Annex A (informative) maps the clause 6 life cycle to the OECD AI system life cycle definition.

Notable quotes

• “engineered system that generates outputs such as content, forecasts, recommendations or decisions for a given set of human-defined objectives” (3.1.4, AI system definition)
• “AI systems do not understand; they need human design choices, engineering and oversight.” (7.1)
• “A key point is that AI outputs are error prone. The output has a probability of being a correct, rather than being absolutely true.” (7.4.1)
• “this document does not use the popular term autonomous to describe automation” (5.13, Table 1 NOTE)
• “in the context of AI, instead of bias, the term unfairness is used to refer to unjustified differential treatment that preferentially benefits certain groups more than others” (5.15.9)
• “Knowledge in the AI domain does not imply a cognitive capability… In particular, knowledge does not imply the cognitive act of understanding.” (3.1.21, Note 1)

What’s speculative vs. asserted

• Asserted (normative definitions): all clause 3 terms and abbreviations.
• Asserted (descriptive): concept explanations in clauses 5–10; the life cycle model is explicitly an example (“This document does not prescribe a specific life cycle model”, 6.1).
• Hedged by the source: feasibility of general AI (“It is not yet known whether ‘general’ AI systems will be technically feasible in the future”, 5.2); trustworthiness/transparency described as topics of ongoing research (5.15.8); driverless-vehicle ubiquity is “anticipated” (10.3).
• Informative only: Annex A (OECD mapping), bibliography.

Topics this feeds

• ISO 22989 — AI Concepts and Terminology — the study Topic page synthesising this standard.
• ISO 42001 — supplies the concept/terminology base for the AIMS standard.
• PECB ISO 42001 Lead Auditor Certification — Phase 1 study text in the 42001 LA pathway.

Open questions raised

• The standard punts specific life cycle prescription to organisations (6.1) and points risk management wholly to ISO/IEC 23894 — clause-level integration appears only when 23894 is ingested.
• Terminology currency: the 2022 text predates the generative-AI wave; it defines no generative-AI terms. (An iso.org status check on 2026-08-17 found Amendment 1 “Generative AI” at draft/ballot stage and Amendment 2 registered — web check, not yet an ingested source.)