ISO/IEC 5259

ISO/IEC 5259 is a six-part series of international standards titled Artificial intelligence — Data quality for analytics and machine learning (ML). Its purpose is to give organisations tools and methods to assess and improve the quality of the data they use to train and evaluate machine learning models.

It was developed by ISO/IEC JTC 1/SC 42 and published from July 2024 onwards. It matters because data quality is not a hygiene issue in machine learning — it is the input that determines whether the model works at all.

Part 1, overview, terminology and examples. Part 2, data quality measures. Part 3, data quality management requirements and guidelines. Part 4, data quality process framework. Part 5, data quality governance framework. Part 6, a Technical Report describing a visualization framework for data quality.

They do genuinely different jobs. Part 1 gives you the language, Part 2 the measuring instruments, Part 3 the obligations, Part 4 the process, Part 5 the board-level view, and Part 6 a way of showing the results.

Only Part 3. ISO/IEC 5259-3 specifies requirements for a data quality management system, and it is the only part in the series written in the language of obligation. Everything else provides a model, a framework, measures or guidance.

This is the single most useful thing to know about the series, and the reason buying Part 2 on its own leaves people with a measuring instrument and nothing to measure against.

Parts 1, 3 and 4 were published in July 2024, Part 2 in November 2024, Part 5 in 2025, and the Technical Report forming Part 6 in 2026. The project was registered in the SC 42 programme in July 2020, so the first parts took four years.

There is no accredited certification scheme for ISO/IEC 5259 in the way there is for ISO/IEC 42001, which has ISO/IEC 42006 governing its certification bodies. What Part 3 does provide is a data quality management system that gives transparency and auditability — the standard is explicit that this can be through self-assessment or through third-party assessment.

In practice 5259 shows up inside a 42001 audit rather than as a certificate of its own. If someone offers to certify you against 5259, ask them precisely what they are accredited to do.

No. This is a reasonable assumption and it describes the wrong deliverable. A Technical Report is informative by design — under ISO's drafting rules a TR cannot contain requirements at all — and it is not on a path to becoming an International Standard.

The deliverable that is provisional is a Technical Specification. ISO publishes a TS where a subject is still under technical development, or where agreement on an International Standard is a future but not immediate possibility; it is issued for use and for feedback, with the aim of eventually being transformed and republished as an International Standard. No part of ISO/IEC 5259 is a TS. Part 6 is a permanent, informative companion to the series.

A Technical Specification addresses work still under technical development, or where agreement on an International Standard is a future but not immediate possibility. It is published for immediate use and to gather feedback, and the intention is that it will eventually be transformed and republished as an International Standard. It is normative — it can contain requirements.

A Technical Report contains information of a different kind altogether: survey data, an informative account, or the state of the art. It is informative and cannot contain requirements, and it is not a step towards anything. TS means not yet. TR means never intended to be.

60.60 means the document has been published and is current. ISO tracks every project through numbered stages: 10 proposal, 20 preparatory, 30 committee, 40 enquiry, 50 approval, 60 publication, 90 review and 95 withdrawal. The second number is the substage.

The codes are public and they are the fastest way to check whether a standard someone is citing actually exists yet. ISO/IEC 5259-1 was approved as a project on 31 July 2020, reached committee draft in August 2022, went to a twelve-week DIS ballot in June 2023, and hit 60.60 in July 2024 — four years, which is normal.

Directly and explicitly. ISO/IEC 5259-1 states that its data life cycle model is derived from ISO/IEC 8183, and Part 5 repeats the point. Part 3 then refines that same life cycle for data quality management purposes.

In other words, 8183 defines the stages that data pass through in an AI system, and 5259 describes how to manage quality across those stages. Reading 5259 without understanding 8183's life cycle is reading the second half of a sentence.

ISO/IEC 42001 is the certifiable AI management system standard, and its Annex A.7 covers data for AI systems. An auditor testing those controls is asking whether your data are managed, provenanced and fit for purpose — which is exactly what ISO/IEC 5259-3 sets out requirements for. 5259-3 lists 42001 in its bibliography.

The practical relationship: 42001 asks whether you manage data quality; 5259 is how you do it and evidence it.

The designation tells you the document's status. In the UK, BSI publishes Parts 1 to 4 as BS EN ISO/IEC 5259, meaning they were taken over as European Standards. Part 5 is published as BS ISO/IEC 5259-5 — a British Standard implementing the international text, without the European designation. Part 6 is published as a PD, a Published Document, because it is a Technical Report rather than a standard.

If an obligation in a contract or a regulatory framework is worded around European Standards, that distinction decides which parts of the series it actually reaches.

Part 2 defines a data quality model as a set of data quality characteristics — categories such as accuracy, completeness or precision — and a data quality measure as a variable whose value is the result of measuring one of them. Measures are how you test whether data meet a requirement, and how you monitor and report quality over time.

The characteristic the series returns to most is representativeness, because training data that does not represent the production population produces a model that is confidently wrong about the groups it saw least of.

Yes. The requirements and recommendations in Part 3 are generic and apply to organisations of every type, size and sector, and Part 4 says the same. There is no SME carve-out and no risk tier to sit below. What varies is how much of the series you need, not whether it applies to you.

ISO/IEC JTC 1/SC 42, the joint ISO and IEC subcommittee for artificial intelligence. Experts are nominated through their national standards bodies and serve in a personal and expert capacity; every participating country votes through its own mirror committee, which in the United Kingdom is BSI's ART/1.

From ISO, from the IEC, or from your national standards body — in the UK, BSI. Note that the parts are sold separately and the series is not cheap, which is precisely why knowing which parts you actually need is worth doing before you buy. This site does not reproduce any of them; they are copyrighted documents.