Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
ISO/IEC 26134:2023 — Artificial Intelligence — Terminology
ISO/IEC 26134:2023
ISO/IEC 26134:2023 establishes a unified and authoritative terminology for the field of artificial intelligence. As AI technologies proliferate across industries — from healthcare diagnostics and autonomous vehicles to natural language processing and predictive analytics — the need for a shared vocabulary has become paramount. This standard provides clear, consistent definitions for over 200 key AI terms, serving as the linguistic foundation for all subsequent AI standards including ISO/IEC 26135 (risk management), ISO/IEC 26136 (life cycle processes), and ISO/IEC 26137 (validation). Without standardized terminology, cross-disciplinary communication, regulatory compliance, and international collaboration in AI development would be fraught with ambiguity and misinterpretation.
Terminology standards are the bedrock of all technical standardization. Before implementing AI risk management (26135) or validation (26137), ensure your team shares a common understanding of terms like ‘AI system’, ‘ML model’, ‘training data’, ‘bias’, and ‘fairness’ as defined in 26134.
Be aware that some terms in this standard differ subtly from their usage in other fields. For example, ‘validation’ in AI context has a different scope than in traditional software engineering. Cross-reference definitions carefully.
The standard was published in 2023 and reflects the state of AI terminology at that time. With the rapid evolution of generative AI and large language models, some emerging terms (e.g., ‘foundation model’, ‘RLHF’) may not yet be included. Stay current with revision activities.
Adopting a common terminology across your organization reduces integration friction, improves audit trail clarity, and simplifies compliance demonstrations. Consider mapping your internal glossary to this standard.
1. Scope and Structure of AI Terminology
The standard categorizes AI terminology into several domains: foundational concepts (e.g., artificial intelligence, machine learning, deep learning, neural network), system properties (e.g., robustness, explainability, transparency, bias, fairness), data-related terms (e.g., training data, validation data, test data, labeling, data quality), process-related terms (e.g., training, inference, fine-tuning, transfer learning), and governance terms (e.g., AI lifecycle, AI stewardship, risk, harm, trustworthiness). Each definition includes not only a concise textual description but also notes and examples that clarify usage in context, making the standard accessible to both technical practitioners and non-specialist stakeholders.
Pay special attention to the definitions of ‘bias’ and ‘fairness’ — these terms have specific technical meanings in the AI context that differ from everyday usage. Misunderstanding these can lead to incorrect compliance assessments.
The distinction between ‘validation’ (ensuring the right system was built) and ‘verification’ (ensuring the system was built right) follows the classic V-model but with AI-specific nuances around data and model behavior.
For systems employing machine learning, the term ‘training’ has a precise statistical meaning. Do not confuse it with ‘learning’ in the human sense — the standard provides the necessary technical precision.
The term ‘AI system’ is defined broadly to include both rule-based and learning-based approaches. This ensures the standard remains relevant as AI technologies evolve beyond current paradigms.
2. Key Definitions and Their Practical Implications
Several definitions in ISO/IEC 26134 carry significant practical implications for AI system design and deployment. The term “trustworthiness” is defined as the ability of an AI system to meet stakeholders’ expectations of reliability, availability, resilience, safety, security, and privacy — encapsulating multiple quality attributes into a single overarching concept. “Explainability” is defined as the ability to provide understandable reasons for AI system outputs, a critical requirement for regulated industries such as finance, healthcare, and autonomous transportation. The standard also provides precise definitions for different types of learning (supervised, unsupervised, semi-supervised, reinforcement learning), enabling clear communication about which approach is being employed.
| Term | Definition (abbreviated) | Practical Significance | Related Concepts |
|---|---|---|---|
| AI System | Engineered system using AI techniques to generate outputs | Scope definition for all AI standards | Machine learning, rule-based system |
| Trustworthiness | Ability to meet stakeholder expectations of reliability, safety, etc. | Holistic quality framework | Robustness, explainability, fairness |
| Bias | Systematic difference in treatment or representation | Compliance and ethics requirement | Fairness, discrimination, equity |
| Explainability | Ability to provide understandable reasons for outputs | Regulatory requirement (EU AI Act) | Transparency, interpretability |
| Training Data | Data used to train an ML model | Data quality governance | Validation data, test data, labeling |
| AI Lifecycle | Evolution of an AI system from conception to retirement | Process management framework | Validation, monitoring, retirement |
The standard’s definition of “AI lifecycle” is particularly important for engineers. It encompasses not only the development phase but also deployment, operation, monitoring, and eventual retirement. This lifecycle perspective ensures that AI governance is not limited to the design phase but extends throughout the entire operational lifetime of the system. The definitions of “data quality” and “data governance” provide the vocabulary needed to implement robust data management practices that are critical for AI system performance and regulatory compliance.
Warning: Organizations frequently misuse the term ‘AI’ to describe conventional statistical models or simple rule-based systems. ISO/IEC 26134 provides a clear definition that distinguishes AI systems from traditional software. Mislabeling can lead to incorrect regulatory classification and inappropriate risk management approaches. Ensure your compliance team understands these distinctions before engaging with regulators.
3. Impact on the AI Standards Ecosystem
ISO/IEC 26134 serves as the terminological foundation for the entire ISO/IEC AI standards family. Its definitions are referenced normatively by ISO/IEC 26135 (risk management), ISO/IEC 26136 (life cycle processes), and ISO/IEC 26137 (validation), among others. This creates a consistent conceptual framework that enables interoperability between different standards and facilitates integrated implementation. For organizations adopting multiple AI standards, starting with terminology alignment reduces confusion, streamlines training, and ensures consistent interpretation of requirements across teams. The standard also supports regulatory compliance efforts by providing terminology that aligns with emerging AI regulations such as the EU AI Act.
When implementing multiple AI standards, use 26134 as the training foundation. Conduct terminology workshops to ensure all team members have a shared understanding of key concepts before diving into risk management or validation processes.
The standard’s definitions of ‘harm’ and ‘risk’ are designed to align with ISO 31000 (risk management), facilitating integrated risk management approaches. This alignment is intentional and beneficial.
Mapping your organization’s internal AI vocabulary to 26134 terms can reveal inconsistencies and gaps in understanding. This is a valuable exercise even if full compliance with the standard is not required.
The standard is expected to be updated as AI technology evolves. Monitor the revision cycle — especially for terms related to generative AI, foundation models, and agentic AI systems.
Frequently Asked Questions
Q: How does ISO/IEC 26134 relate to the EU AI Act terminology?
A: The EU AI Act uses its own set of definitions, but there is significant alignment between the Act’s terminology and ISO/IEC 26134. Organizations compliance with both frameworks benefit from using 26134 as their internal terminology baseline, then mapping to the Act’s specific definitions where differences exist.
A: The EU AI Act uses its own set of definitions, but there is significant alignment between the Act’s terminology and ISO/IEC 26134. Organizations compliance with both frameworks benefit from using 26134 as their internal terminology baseline, then mapping to the Act’s specific definitions where differences exist.
Q: Is ISO/IEC 26134 applicable to non-ML AI systems?
A: Yes. The standard defines AI broadly to include both rule-based and learning-based approaches. Definitions for ‘expert system’, ‘knowledge representation’, and ‘symbolic AI’ are included alongside machine learning terminology.
A: Yes. The standard defines AI broadly to include both rule-based and learning-based approaches. Definitions for ‘expert system’, ‘knowledge representation’, and ‘symbolic AI’ are included alongside machine learning terminology.
Q: Can I use this standard as a glossary for AI procurement contracts?
A: Absolutely. Referencing ISO/IEC 26134 definitions in procurement contracts is a best practice to ensure alignment between buyers and suppliers on key AI terminology. This reduces ambiguity in statements of work, acceptance criteria, and service level agreements.
A: Absolutely. Referencing ISO/IEC 26134 definitions in procurement contracts is a best practice to ensure alignment between buyers and suppliers on key AI terminology. This reduces ambiguity in statements of work, acceptance criteria, and service level agreements.
Q: How often is this standard updated?
A: The standard was published in 2023. ISO standards typically undergo periodic review every 3-5 years. Given the rapid pace of AI evolution, earlier revision is possible.
A: The standard was published in 2023. ISO standards typically undergo periodic review every 3-5 years. Given the rapid pace of AI evolution, earlier revision is possible.
Q: Does this standard define ‘Artificial General Intelligence’ (AGI)?
A: The standard focuses on current AI technologies and does not include a definition of AGI, as AGI remains a theoretical concept without concrete implementation. The scope is limited to contemporary AI systems.
A: The standard focuses on current AI technologies and does not include a definition of AGI, as AGI remains a theoretical concept without concrete implementation. The scope is limited to contemporary AI systems.
