ISO/IEC 22989 is the international standard that fixes the vocabulary of artificial intelligence: what an AI system is, what training data is, what bias, transparency and risk mean when engineers, auditors and lawyers sit at the same table. Published in July 2022, it is the text ISO/IEC 42001 points to for its terms, and it is free to download. It is also widely misread. Its definitions do not match the EU AI Act word for word, and the 2022 edition was written before generative AI reached the enterprise. We read the full text and counted what is in it. This guide covers what ISO/IEC 22989 contains, where it parts ways with the law, and how to use it without importing its blind spots.
Key takeaways
- ISO/IEC 22989 is a terminology and concepts standard from ISO/IEC JTC 1/SC 42, first edition July 2022, adopted in Europe as EN ISO/IEC 22989:2023.
- By our count it holds 117 numbered term entries in seven groups, plus concept clauses on trustworthiness, the AI system life cycle and stakeholder roles.
- It cannot be certified. It contains definitions and no requirements, and it sits in ISO’s list of publicly available standards at no charge.
- “Provider”, “autonomy” and “risk” do not mean the same thing in the standard and in the EU AI Act.
- The 2022 text has no generative AI term. Amendment 1 on generative AI is in the approval phase, Amendment 2 is a committee draft.
What is ISO/IEC 22989?
Its full title is Information technology, Artificial intelligence, Artificial intelligence concepts and terminology. It was prepared by ISO/IEC JTC 1/SC 42, the joint ISO and IEC subcommittee that also produced the AI management system standard. The catalogue entry states the purpose in two sentences: the document establishes terminology for AI and describes concepts in the field, and it can be used in the development of other standards and to support communication among diverse stakeholders. It applies to every type of organisation, from a commercial enterprise to a government agency or a non-profit. That last point explains its position. ISO/IEC 22989 is the base layer of the SC 42 family. A management system standard, a risk guidance document and a machine learning framework can only be read together if “model”, “training” and “validation” mean one thing across all of them.
A dictionary, not a requirement
Clause 2 is a single sentence: there are no normative references in the document. The rest follows the same logic. The text defines and describes. It never says what an organisation shall do, so there is nothing to audit against and no certificate to obtain. Training providers sell course certificates for individuals, which is a different product. Two practical consequences follow. First, a vendor that claims to be “ISO/IEC 22989 compliant” is claiming something no scheme can verify. The honest version of that sentence is “we use its vocabulary”. Second, there is no cost barrier. Since September 2023 ISO/IEC 22989 has been listed among ISO’s Publicly Available Standards, which can be downloaded free of charge. A corrected version of the 2022 text was issued in December 2025, so check the date of the copy on your shared drive.
What ISO/IEC 22989 contains
Clause 3 carries the definitions. We counted the numbered entries group by group. <table header-row=”true”> <tr> <td>Term group</td> <td>Entries</td> <td>Examples</td> </tr> <tr> <td>Terms related to AI</td> <td>35</td> <td>AI agent, AI system, autonomy, heteronomy, general AI, narrow AI, model, inference</td> </tr> <tr> <td>Terms related to data</td> <td>15</td> <td>dataset, ground truth, test data, validation data, production data, data annotation</td> </tr> <tr> <td>Terms related to machine learning</td> <td>17</td> <td>machine learning, trained model, hyperparameter, retraining, reinforcement learning, human-machine teaming</td> </tr> <tr> <td>Terms related to neural networks</td> <td>10</td> <td>neural network, deep learning, convolutional neural network, long short-term memory</td> </tr> <tr> <td>Terms related to trustworthiness</td> <td>18</td> <td>accountability, bias, controllability, explainability, predictability, reliability, resilience, risk, transparency</td> </tr> <tr> <td>Terms related to natural language processing</td> <td>18</td> <td>natural language understanding, machine translation, sentiment analysis, emotion recognition</td> </tr> <tr> <td>Terms related to computer vision</td> <td>4</td> <td>computer vision, image recognition</td> </tr> </table> That makes 117 entries. The trustworthiness group is the one compliance teams reuse most, because it gives a neutral wording for properties that appear in every AI policy. Bias, for instance, is described as a systematic difference in the treatment of certain objects, people or groups compared with others. That is a descriptive definition, not a synonym for unlawful discrimination, and the distinction matters when a data scientist and a lawyer discuss the same test result. Trustworthiness itself is the ability to meet stakeholder expectations in a verifiable way, which is a good one-line brief for an accountability programme. Clauses 5 to 9 then describe concepts rather than define words: AI concepts, the AI system life cycle, a functional overview of an AI system, the AI ecosystem, and fields of AI such as computer vision and planning. Two parts of that material are directly usable in a governance programme. The first is the life cycle. ISO/IEC 22989 says it does not prescribe a model, then offers eight stages as an example:
- inception
- design and development
- verification and validation
- deployment
- operation and monitoring
- continuous validation
- re-evaluation
- retirement
An informative Annex A maps these stages to the OECD’s description of the AI system life cycle. The second is clause 5.19 on stakeholder roles. It names six: the AI provider, which supplies products or services that use AI systems; the AI producer, which designs, develops, tests and deploys them, with the AI developer as a sub-role; the AI customer, which uses an AI product or service, with the AI user as a sub-role; the AI partner, a family that covers the system integrator, the data provider, the AI auditor and the AI evaluator; the AI subject, meaning anyone impacted, data subjects included; and the relevant authorities, split into policy makers and regulators. A note adds that one organisation can hold several roles at once.
ISO/IEC 22989 vs the EU AI Act: same words, different meanings
The drafters of Regulation (EU) 2024/1689 did not take their central definition from ISO. They took it from the OECD, as revised in November 2023. The result is two vocabularies that overlap on the page and diverge in meaning. Five terms cause most of the damage. <table header-row=”true”> <tr> <td>Term</td> <td>In the standard</td> <td>In the EU AI Act</td> <td>Consequence</td> </tr> <tr> <td>AI system</td> <td>”engineered system that generates outputs such as content, forecasts, recommendations or decisions for a given set of human-defined objectives”</td> <td>A machine-based system that operates with varying levels of autonomy, may adapt after deployment and infers, for explicit or implicit objectives, how to generate outputs (Article 3(1))</td> <td>Inference, autonomy and adaptiveness are legal criteria. The ISO wording cannot decide whether the Act applies</td> </tr> <tr> <td>Provider</td> <td>An organisation that supplies products or services using AI systems</td> <td>Whoever develops an AI system, or has one developed, and places it on the market or puts it into service under its own name (Article 3(3))</td> <td>The legal provider is closer to the ISO “AI producer”</td> </tr> <tr> <td>User / deployer</td> <td>AI customer and AI user: an organisation that uses an AI product or service</td> <td>Deployer: uses an AI system under its authority, outside personal non-professional activity (Article 3(4))</td> <td>”Deployer” does not appear anywhere in the 2022 text</td> </tr> <tr> <td>Autonomy</td> <td>The capacity of a system to modify its intended domain of use or its goal without external intervention, control or oversight</td> <td>Some degree of independence of action from human involvement</td> <td>Under ISO vocabulary most production systems are automated, not autonomous. Under the Act they have autonomy</td> </tr> <tr> <td>Risk</td> <td>”effect of uncertainty on objectives”, positive or negative</td> <td>The combination of the probability of an occurrence of harm and the severity of that harm (Article 3(2))</td> <td>An ISO risk register can hold opportunities. An AI Act risk assessment cannot</td> </tr> </table> The most expensive confusion is the first role. An organisation that sees itself as an “AI customer” in ISO/IEC 22989 terms, because it commissioned a system from an integrator, becomes a provider under Article 3(3) the day it puts that system into service under its own name. Every high-risk obligation, from the technical file to the conformity assessment, then lands on it and not on the integrator. A role map written only in ISO vocabulary hides that outcome. The two texts also answer different questions. ISO/IEC 22989 describes a field so that people can talk about it. The law allocates obligations and penalties. For scoping, the legal definition always wins, and the Commission’s guidelines on the AI system definition of 6 February 2025 are the reference for borderline cases. The same holds for the autonomy row: the ISO definition is useful when you design human oversight, because it separates automation from genuine goal-changing autonomy, but it must not be used to argue that a system falls outside the Act.
What the 2022 text does not define: generative AI and the amendments
We searched the 2022 edition for the vocabulary that dominates AI governance in 2026. The following words appear zero times: “generative”, “foundation model”, “large language model”, “prompt”, “hallucination” and “fine-tuning”. The legal vocabulary is absent as well: “deployer”, “importer”, “distributor”, “intended purpose”, “high-risk” and “substantial modification” are nowhere in the text. Two near misses deserve a warning. “General AI” in ISO/IEC 22989 means artificial general intelligence, a system that handles a broad range of tasks. It has nothing to do with the general-purpose AI model of the AI Act. And the standard does define an “AI agent”, as an automated entity that senses and responds to its environment and takes actions to achieve its goals, but that sentence was written before LLM-based agents that call tools and act across enterprise systems. SC 42 is closing the gap through amendments:
- Amendment 1, Generative AI. A draft amendment circulated in 2025, with a parallel European enquiry. When we checked on 8 October 2026, the ISO catalogue listed it as a final draft in the approval phase.
- Amendment 2. A committee draft at ISO. The European project was registered on 28 October 2025 under CEN/CLC/JTC 21, according to the DIN project page.
- Part 2. A work item for a second part dedicated to healthcare has been opened.
Until Amendment 1 is published, generative AI terms have to come from elsewhere: Article 3(63) of the AI Act for the general-purpose AI model, and NIST AI 600-1 for the risk vocabulary. Whatever you choose, date your glossary. A definition of “hallucination” written in 2024 will have to be reconciled with the amendment when it lands.
Where ISO/IEC 22989 sits in the standards stack
ISO/IEC 22989 matters because of what refers to it. ISO/IEC 42001:2023, the certifiable AI management system standard, lists it as a normative reference and takes its terms from it. An auditor reading your AI policy during a 42001 audit will interpret “AI system” and “risk” through this vocabulary unless your documentation says otherwise. ISO/IEC 23894 on AI risk management and ISO/IEC 23053 on machine learning systems build on the same terms. The influence extends beyond ISO. The NIST AI RMF 1.0 marks its own description of an AI system as adapted from the OECD Recommendation and from ISO/IEC 22989:2022, which is one reason the NIST framework and the ISO family translate into each other with little friction. In Europe the text was adopted as EN ISO/IEC 22989:2023, published on 28 June 2023, so national standards bodies carry it as a national standard. The harmonised standards requested for the AI Act are being drafted by CEN-CENELEC JTC 21. According to public trackers, none had been cited in the Official Journal by July 2026, so no standard yet gives the presumption of conformity of Article 40. A vocabulary standard would not give that presumption for any requirement in any case. That is the central message of our guide to the AI Act standards stack: terminology aligns people, it does not discharge duties. The calendar leaves time to get the vocabulary right. After the Digital Omnibus on AI, Regulation (EU) 2026/1744, the obligations for Annex III high-risk systems apply from 2 December 2027 and those for Annex I product-embedded systems from 2 August 2028. Transparency duties under Article 50 have applied since 2 August 2026.
How to use ISO/IEC 22989: a five-step glossary method
A shared glossary is the cheapest control in AI governance. Here is how to build one that survives both an ISO audit and a regulator’s question.
- Adopt the standard as the default glossary. State in the AI policy that terms carry the meaning given in ISO/IEC 22989 unless the deviation register says otherwise. That single sentence removes most definitional debates from project meetings.
- Keep a deviation register. For every term the applicable law defines differently, record two columns, the standard and the law, and a rule that says which one governs which document. Start with the five rows of the table above. Legal scoping, contracts with EU exposure and regulatory filings follow the law. Engineering documentation can follow the standard.
- Record roles twice for each AI system. In the AI inventory, store the ISO/IEC 22989 clause 5.19 role, which is useful for contracts and supply chain mapping, next to the AI Act role, which decides the obligations. When the two diverge, as in the commissioned-system case, the record shows it at once.
- Align life cycle vocabulary. Map your internal stage names to the eight stages of ISO/IEC 22989, so that evidence from data science, IT and compliance is filed against the same stage names.
- Date the glossary and review it on triggers. The triggers are known: publication of Amendment 1, citation of a harmonised standard on terminology, a new Commission guideline that changes how a definition is read.
An AI governance platform such as AI Sigil keeps the glossary, the roles and the records in one AI registry, so that a change of definition is applied once and not in twelve spreadsheets.
FAQ
What is ISO/IEC 22989 in simple terms? It is the dictionary of the AI standards family. Published in July 2022 by the joint ISO and IEC subcommittee on artificial intelligence, it defines 117 terms by our count, from “AI system” to “training data” and “explainability”, and describes concepts such as the AI system life cycle and stakeholder roles. Other standards, including the AI management system standard, rely on it so that the same word means the same thing from one document to the next. Is ISO/IEC 22989 free? Yes. ISO/IEC 22989 has been listed in ISO’s Publicly Available Standards since September 2023 and can be downloaded at no charge from the official ISO site after accepting the licence terms. National adoptions sold by standards bodies carry a price, but the content is the same. Avoid unofficial copies: they are often outdated, and a corrected version of the 2022 text was issued in December 2025. Can a company be certified against ISO/IEC 22989? No. The document contains definitions and descriptions, not requirements, so there is nothing for a certification body to audit. The certifiable standard in the family is ISO/IEC 42001, which uses this vocabulary. Courses on the terminology standard deliver certificates of attendance or competence for individuals, which say nothing about an organisation. Treat any claim of company-level compliance with it as marketing. How does ISO/IEC 22989 define an AI system? As an engineered system that generates outputs such as content, forecasts, recommendations or decisions for a given set of human-defined objectives. The definition is short and technology-neutral. It does not mention inference, autonomy or adaptiveness, which are the criteria the EU AI Act uses in Article 3(1). For that reason it cannot be used on its own to decide whether a system falls under the Act. How does ISO/IEC 22989 relate to ISO/IEC 42001? ISO/IEC 42001 lists the terminology standard as a normative reference, which means its terms apply when the management system standard is read and audited. In practice an organisation preparing for 42001 certification should use the same definitions in its AI policy, its risk assessment and its AI system inventory, and document any deliberate deviation, for example where a legal definition must prevail. Does ISO/IEC 22989 cover generative AI? Not in the 2022 edition. The words “generative”, “foundation model”, “large language model”, “prompt” and “hallucination” do not appear in it. An amendment dedicated to generative AI is in the approval phase at ISO as of October 2026, and a second amendment is at committee draft stage. Until publication, organisations take their generative AI vocabulary from the AI Act and from NIST AI 600-1.
Conclusion
Most disputes in AI governance start as vocabulary problems: two teams using “provider”, “risk” or “autonomy” with different meanings and discovering it during an audit. ISO/IEC 22989 removes a large share of that friction, at no cost, and it is the vocabulary an ISO/IEC 42001 auditor will assume. Its limits are just as clear. It is a 2022 text with no generative AI term, it sets no requirement, and its definitions are not those of the EU AI Act. Use it for what it is: adopt it as the default glossary, record where the law says otherwise, write down both roles for each system, and date the result. When Amendment 1 is published, the review will take an afternoon instead of a quarter.