The draft defines a generative model as one that produces new data from a learned distribution, and lists the model families that fall under it.
What the base framework establishes
The public summary says the amendment adds a generative model definition at 6.5.3.2.10. A generative model is described as producing new data from a distribution it has learned, and the definition is said to cover both probabilistic and non-probabilistic examples. The listed families include autoencoders, variational autoencoders, generative adversarial networks, normalising flows, transformers, diffusion models, Boltzmann machines, restricted Boltzmann machines, Hopfield networks and autoregressive models. The final wording is not held. The practical effect of the public record is that the framework now names the family and gives a buyer a defensible way to say whether a system is generative rather than merely predictive.
Why it matters to an AI management system
Buyers argue about whether a system is generative, and the answer changes which risks and which regulation apply. A learned-distribution definition is a test a governance team can apply: does the system produce new samples from a distribution, or does it assign a label or a value? The list of families also stops the conversation from collapsing into large language models alone.
The public draft adds material at 6.5.3.2.10. Its final wording is not held, so this page records what the draft summary states and does not present it as published text. New in the draft. The public summary lists the definition and the families; the final wording is not held and is not stated.
Evidence a reviewer samples
| Category | Artefact |
|---|---|
| Generative classification note | Why each system is or is not generative under the learned-distribution test. |
| Family record | The model family and the output it produces. |
| Output handling note | How generated content is labelled, reviewed and retained. |
One question to prepare for. Does this system produce new samples, or assign a label or value to an input?
Common gaps
- Every language model is called generative without checking what the system actually outputs.
- A generative system is classified as predictive because its output is used in a decision.
- Generated content is stored with no label and no review record.
Prepare before publication
Classify each system against the learned-distribution test and record the family. Add an output handling note for every generative system, and give engineering a short briefing on the new definitions so the description uses them consistently.
Dates that matter
No separate date attaches to 6.5.3.2.10; the record dates are the stage date, the planning estimate and the CEN expectation stated above.
Cross-references
6.2 Task · 6.5 Tools · 6.5.3.2.9 Transformer architecture and algorithm · 7.5 Self-supervised machine learning
Source
Source for 6.5.3.2.10: ISO/IEC 23053:2022/FDAmd 1, new clause 6.5.3.2.10 (public draft summary). The public sample stops before 6.5.3.2.10, so this page relies on the public contents and the public summary and does not quote the clause. The two official records are linked on this page so any line can be checked.
Not legal advice; does not certify. A reading of the framework and the evidence a reviewer commonly samples, not a substitute for the standard or your auditor's rules.