01The short answer
Three things are true at once.
The framework is published
ISO/IEC 23053:2022 describes the components of an ML system and the pipeline that builds one. It is a vocabulary, not a rulebook.
The amendment is not published
Amendment 1 is at stage 50.00. The public draft summary names its additions; the final text is not held and is not invented here.
No certification exists
You cannot be certified to 23053, and no transition audit follows from it. ISO/IEC 42001 is the certifiable management system standard.
The buying question is not "how do we certify". It is "can we show, before the publication date, that our generative systems are described in the current vocabulary and that their risks are recorded".
02Free check
Your position, read against the record.
The check uses your role, your systems, your 42001 position, a date and one data question. It returns a named planning position and never defaults to a pass. Your answers are computed in the browser and are not sent to us.
03The rules this is built from
Read from the real rules, not invented.
The base record is written from ISO/IEC 23053:2022: 21 substantive units, read from the public contents and the public sample at clauses 1 to 6.3. The amendment record carries the 3 additions the public draft summary states. Every page names its unit and links the ISO record, so a reader can check any line.
The base framework
21 units, published edition, 36 pages, ISO/IEC JTC 1/SC 42. Open the framework record.
The amendment
3 public draft additions at stage 50.00. Open the amendment record.
Nothing here is guessed while you wait. Every answer is written from the public record and points to the exact rule it came from.
04What the draft adds
Two new entries, one expanded note, one reference.
- A new entry at 6.5.3.2.9 describes transformer architecture and algorithm, including encoder-decoder structure, self-attention and the difference between encoder-only and decoder-only use.
- A new entry at 6.5.3.2.10 defines a generative model as one that produces new data from a learned distribution, and lists the model families it covers.
- The note at 7.5 on self-supervised learning is expanded with the model collapse risk of training on generated output over successive cycles.
- The bibliography is updated with a survey of AI-generated content.
The public summary is the only account of the draft's technical content. It is a summary, not the text, and this site says so on every page that relies on it. Read the full amendment record.
What it is
An extension of the framework's model vocabulary at the point where a practitioner chooses an algorithm family. The base structure, the pipeline and the annex are not replaced.
What it is not
A new management system requirement, a certification scheme, a transition deadline or a legal obligation. Those are separate instruments, and the 42001 record is where they belong.
05The record
One page for every public unit.
32 pages: 28 base clauses and subclauses, 1 informative annex, and 3 pages for the amendment's additions. Each page states what the base framework establishes, whether the draft touches it, the evidence a reviewer samples and the preparation step.
| Unit | Title | Amendment status |
|---|---|---|
1 | Scope | Not touched by the draft |
2 | Normative references | Not touched by the draft |
3.1 | Model development and use | Not touched by the draft |
3.2 | Tools | Not touched by the draft |
3.3 | Data | Not touched by the draft |
4 | Abbreviated terms | Not touched by the draft |
5 | Overview | Not touched by the draft |
6.1 | Overview of the machine learning system | Not touched by the draft |
6.2 | Task | Not touched by the draft |
6.3 | Model | Not touched by the draft |
6.4 | Data | Not touched by the draft |
6.5 | Tools | New material added by the draft |
6.5.3 | Categories of machine learning algorithms | New entries added by the draft |
6.5.3.2.9 | Transformer architecture and algorithm | New in the draft |
6.5.3.2.10 | Generative model | New in the draft |
7.1 | General | Not touched by the draft |
7.2 | Supervised machine learning | Not touched by the draft |
7.3 | Unsupervised machine learning | Not touched by the draft |
7.4 | Semi-supervised machine learning | Not touched by the draft |
7.5 | Self-supervised machine learning | Note expanded in the draft |
7.6 | Reinforcement machine learning | Not touched by the draft |
7.7 | Transfer learning | Not touched by the draft |
8.1 | General | Not touched by the draft |
8.2 | Data acquisition | Not touched by the draft |
8.3 | Data preparation | Not touched by the draft |
8.4 | Modelling | Not touched by the draft |
8.5 | Verification and validation | Not touched by the draft |
8.6 | Model deployment | Not touched by the draft |
8.7 | Operation | Not touched by the draft |
8.8 | Example machine learning process based on the pipeline | Not touched by the draft |
A | Annex A: Example data flow and data use statements | Not touched by the draft |
BIB | Bibliography and the AI-generated content reference | New reference in the draft |
06The paid kit
The register a reviewer can walk.
The kit carries every unit with an owner, a due date, a status and an evidence pointer. It adds the dated plan, the synthetic data rule, the model collapse check and the supplier questions for foundation models. PDF for the meeting, CSV for your register, XLSX for the team.
USD 799 one off. No account. Released against a paid Stripe session.
What the free record already gives you
- All 32 unit pages, cited
- The amendment record and what is not held
- The scope and date reading
- The method, the sources and the changelog
- The dated plan
The kit adds the working register and the columns. It does not add facts that are not on the free record.
07Four markets
The same framework, four different obligations around it.
The framework is international. What a governance lead must actually do is set by the market. The record names the held instruments for the United States, the European Union, the United Kingdom and Australia, and says where a regime is named but not held.
08Questions
What buyers ask first.
Is ISO/IEC 23053 certifiable?
No. ISO/IEC 23053 is a framework and terminology standard. It sets no certifiable requirements, and no certification scheme exists against it. The management system standard that an organisation can be certified to is ISO/IEC 42001.
What is Amendment 1?
A final draft amendment to ISO/IEC 23053:2022, titled Generative AI. The public draft adds a transformer architecture entry and a generative model definition to the algorithm catalogue, expands the note on model collapse at 7.5, and updates the bibliography.
When will it be published?
ISO has not announced a date. The ISO stage is 50.00, recorded on 28 August 2026. This site uses 16 February 2027 as a planning estimate, 120 days from the build date. The CEN-CENELEC adoption project lists 4 April 2027 for the European text.
Does the amendment create a transition deadline?
No. No accreditation body has published a transition rule, and there is nothing to transition a certificate to. The date matters because a 42001 audit after it will expect the AI system description to use the current vocabulary.
Who is this for?
AI governance leads, ISO/IEC 42001 implementers and ML platform leads in organisations that build, buy or operate generative AI. The free check says plainly when a role is outside its scope.
What is in the USD 799 kit?
The full obligation register for all 32 units, the evidence list, the dated plan, the synthetic data rule and the generative system questions, exported as PDF, CSV and XLSX. One payment, no subscription.
What does the USD 49 a month watch do?
It emails you when the ISO stage, the public draft summary, the CEN date, the parallel 22989 amendment or a recorded date changes. The email carries the diff, not a newsletter.
09Who wrote this
Gerard Blokdyk, The Art of Service.
I build compliance records that state each line in plain English and point to the source behind it. This record was written from the public ISO pages, the public base sample and the public draft summary on 21 October 2026. It is not affiliated with ISO or IEC and it does not certify anyone.