Self-supervised learning builds its own supervision from the data, for example by predicting a hidden part of an input. The draft expands a note here on model collapse.
What the base framework establishes
The approach uses the structure of the data as the training signal rather than a separate label. A model can be asked to predict a masked token, the next item in a sequence or a missing part of an input, and the answer is already in the data. This is the approach behind most large language models and many generative systems, which is why the amendment reaches it. The public summary of the draft says the note at 7.5 is expanded to warn that training a model on output from generative AI, self-supervised on synthetic data, can reduce performance over successive iterations, a phenomenon called model collapse that can worsen with repeated cycles.
Why it matters to an AI management system
Model collapse is a governance risk, not a curiosity. If a training pipeline recycles generated content, each generation can lose the tails of the real distribution, and the loss may be invisible in a single evaluation. The record needs a rule about synthetic data and a check across generations, or the organisation cannot show that its model still reflects the world.
The public draft expands the note at 7.5. The base note is not in the public sample in full, and the expanded wording is not held. The draft expands the existing note at 7.5 on model collapse. The base note is held in the public record only as a summary; the final expanded wording is not held.
Evidence a reviewer samples
| Category | Artefact |
|---|---|
| Synthetic data rule | The policy on using generated content in training and fine-tuning. |
| Generation check | A comparison of model quality across data generations or training cycles. |
| Training data provenance | The source, date and share of synthetic material in each training data. |
One question to prepare for. How much of the training data is model-generated, and has quality been checked across cycles?
Common gaps
- Generated content is mixed into training data with no share recorded.
- Quality is checked once, at the first generation, and never again.
- The supplier's synthetic data practice is unknown and not asked about.
Prepare before publication
Write a synthetic data rule with a share limit and a provenance field. Add a generation check to the retraining checklist and ask each model supplier for its practice.
Dates that matter
No separate date attaches to 7.5; the record dates are the stage date, the planning estimate and the CEN expectation stated above.
Cross-references
6.5.3.2.10 Generative model · 7.1 General · 8.2 Data acquisition
Source
Source for 7.5: ISO/IEC 23053:2022, clause 7.5, note expanded by FDAmd 1 (public draft summary). The public sample stops before 7.5, 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.