Overview
The Draft International Standard Amendment 1 (DAmd 1) to ISO/IEC 22989:2022, titled Information technology — Artificial intelligence — Artificial intelligence concepts and terminology — Amendment 1: Generative AI, has entered its DIS (Draft International Standard) phase as of August 2025. This amendment has been launched for a 12‑week ballot by ISO members, marking a significant step in its formalisation. The document seeks to extend the foundational AI terminology set out in ISO/IEC 22989:2022 by integrating terminology relevant to generative AI concepts.
Who is it relevant to?
This amendment addresses a broad spectrum of stakeholders across the AI ecosystem. It is particularly pertinent to regulatory authorities, quality and compliance managers, AI governance teams, standards professionals and legal advisors involved in AI policy.
Technology providers developing generative AI models, certification bodies referencing ISO terminology, and organisations aligning with EU-level AI regulations, and looking to ground their systems in internationally recognised concepts, will find this update essential.
Key Requirements
As this document is a terminology amendment, its primary focus is to introduce standardised definitions and conceptual clarity for generative AI. While it does not impose prescriptive obligations, it will, once published, serve as the authoritative lexicon for generative AI concepts. This ensures consistent interpretation across regulatory assessments, conformity demonstrations, technical documentation, and cross-domain communications.
Here are a few new definitions / terminology added:
Selected Definitions from ISO/IEC 22989:2022/DAmd 1 (DIS)
| Clause | Term | Definition | Notes |
|---|---|---|---|
| 3.1.36 | AI model | Model (3.1.23) used in an AI system (3.1.4). | – |
| 3.1.37 | Generative artificial intelligence (Generative AI) | Discipline concerning research and development of mechanisms, methodologies and applications of generative AI systems (3.1.38). | Generative AI is a subdiscipline of artificial intelligence (3.1.3). |
| 3.1.38 | Generative AI system (GenAI system) | AI system (3.1.4) based on techniques and models (3.1.23) that aim to generate new content. | Examples include text, audio, code, video, image. Generated content may be new information or new ways to express pre-existing information. |
| 3.1.39 | Probability distribution | Function that relates all possible outcomes of an observation on a system with the probability of their occurring. | Source: ISO 10303-2:2024, 3.1.88. |
| 3.1.40 | Retrieval-augmented generation system (RAG system) | Generative AI system (3.1.38) that retrieves relevant existing data to better inform the generation of new content. | – |
| 3.1.41 | Token | Unit of content that an AI model (3.1.36) treats as semantically meaningful. | – |
| 3.3.18 | Attention | Mechanism for weighting the importance of different parts of a chunk of input data. | – |
| 3.3.19 | Foundation model | AI model (3.1.36) usable for or adaptable to a wide range of tasks in one or more domains. | Built typically with supervised or self-supervised ML on large data; not limited to generative AI. |
| 3.3.20 | Large language model (LLM) | Machine learning model (3.3.7) that encodes natural language (3.6.7) with many parameters to perform NLP (3.6.9) tasks. | Used for tasks such as generation, summarisation, translation, classification. Requires large datasets and significant compute. |
| 3.3.21 | Self-attention | Attention (3.3.18) where the object to compare belongs to the same set as the compared elements. | – |
| 3.3.22 | Self-supervised machine learning | Machine learning (3.3.5) where supervised learning methods are applied to unlabelled data using implicit labels. | – |
| 3.4.11 | Generative adversarial network (GAN) | Neural network with generators that create synthetic samples and discriminators that distinguish them from real ones. | Generates synthetic data by learning from training data via adversarial testing. |
| 3.4.12 | Transformer (algorithm) | Neural network that models context and structure in sequential data using self-attention. | – |
| 3.4.13 | Transformer (architecture) | Neural network with encoder and decoder, both using the transformer (3.4.12) algorithm. | – |
| 3.4.14 | Variational autoencoder (VAE) | Neural network with encoder and decoder in a probabilistic framework for lower-dimensional representation of input. | By sampling from the learned distribution, a VAE can generate data. |
| 3.4.15 | Diffusion model | Neural network with forward noise-adding process, reverse denoising process, and sampling procedure. | Latent-variable generative model; often used in image generation and manipulation. |
| 3.6.19 | Prompt | Input to a generative AI system that provides instructions on how to process the input. | Prompts may be fixed or editable; can include formatting instructions and output controls. |
Strategic Considerations
Organisations should prepare to adopt and reference these updated terms to enhance clarity and alignment with global regulatory developments. Integration into AI governance frameworks, compliance documentation, process descriptions, and training materials will be essential. Regulatory reviewers and auditors may begin expecting usage aligned with this updated terminology. Early familiarisation with the draft terms—and participating in national body feedback where feasible during the DIS phase—can provide strategic advantage.
Conclusion
The publication of Amendment 1 to ISO/IEC 22989:2022 represents a foundational step in codifying the language of generative AI. Its integration into the ISO lexicon will facilitate more precise discourse across technical, regulatory, and compliance contexts in AI. For regulatory and quality professionals, the amendment underscores the growing necessity of harmonised terminology in aligning policy, oversight, and industry development.
How Blue Arrow Can Help
Blue Arrow, as an AI technology regulatory consultancy and EU Authorised Representative, can support stakeholders at each stage of this development. We help ensure that changes in foundational terminology become an organisational asset rather than a challenge to compliance integrity.
