Authorize reuse of a model for a purpose outside its stated intent
Draft
Authorize reuse of a model for a purpose outside its stated intent
Layer 4. Control AUTAuthorization
Allocation
| L4-AUT-04 | |
|---|---|
| Decides | Business Accountable Executive |
| Consulted | DPO, General Counsel and Model Owner |
| Executes | Model Owner |
| Evidence | Purpose extension record with the original intent stated |
In plain terms
Decide a model validated for one use may serve another. The original validation does not transfer, and the model card describes the first purpose.
What is being judged
Whether the evidence supporting the first purpose supports the second. Usually it partially does, which is what makes this decision easy to skip.
Four things travel badly. Validation data, which reflected the first population. Acceptance thresholds, set for the first purpose’s consequence. The lawfulness basis for the training data, which was determined for the first purpose and does not automatically extend. The classification, since the second purpose may be material where the first was not.
What this decision does not cover
It does not re-authorize the model technically. Where the extension is material, L4-AUT-01 fires again on the new purpose.
When it fires
On event. When a model is applied to a decision, population or process outside its model card’s stated purpose. When another team requests access to an existing model. On integration of a model into a workflow it was not built for.
On cycle. None. Purely event-driven, and the trigger is the hard part.
What you need before deciding
The model card’s stated purpose. What the second purpose actually is, in the same terms. The lawfulness determination behind the training data. The classification of the second use.
How this goes wrong
Reuse by integration: nobody decided to extend the purpose; someone called an existing endpoint from a new workflow. This is the common case and it produces no record at all. The purpose that was never stated: a model card describing capability rather than purpose, so no extension is detectable because no boundary was drawn. Consent drift: the second purpose relies on training data whose lawfulness basis covered only the first, which is the same failure L3-12 addresses for features.
Related decisions
Upstream L4-AUT-01, L3-02 thresholds, L4-REG-03 lawfulness.
Parallel L3-12 feature reuse, the same failure at data level.
Downstream L4-CLS-04 reclassification where the new purpose is material.
Instrument references
ISO/IEC 42001 Annex A addresses intended use documentation. EU AI Act Article 25 addresses substantial modification, which is adjacent but framed around the provider relationship rather than internal reuse.
Correction
The maintainer answers corrections. There is no service level. Responses are best-effort and opportunistic within a reasonable time: a correction raised on a Monday is answered that week or sooner.