Skip to content
Software development
ASD ISMISM-2103

ASD ISM 2103 Get explicit data owner consent before training models on AI application data

What it says

Organisational data generated, collected or processed by AI applications is not used for training, fine-tuning or improving AI models unless informed and explicit consent has been obtained from data owners in advance.

Quoted as published. Everything else on this page is written by Control Stack.

In plain English

Do not use organisational data from AI applications you develop to train, fine-tune or improve AI models unless its data owners gave informed, explicit consent beforehand.

In AI applications the organisation develops, organisational data generated, collected or processed, such as prompts, uploaded files and outputs, must not be used to train, fine-tune or improve AI models without informed and explicit consent obtained in advance from its data owners, those with authority over it. Informed means they knew which data, which models and what purpose. Explicit means a clear yes, not a default setting or a clause in general terms. Any model counts, including a vendor's. Using data only to answer a request is outside this control.

Why it matters

Data used to train or fine-tune a model can be hard to remove and may resurface in the model's outputs, including to a vendor's other customers. Requiring informed, explicit consent in advance keeps that decision with the data owners, not a vendor's default settings.

How to implement

  1. List every way AI application data could reach model training.

    Cover third-party model APIs and AI services, your own fine-tuning pipelines, feedback buttons, vendor review of conversations and analytics exports, and record each one's current data-use setting.

  2. Turn off model training on organisational data by default.

    Use each AI service's opt-out setting or a service tier whose contract excludes training, and record the data types and their ownership in the contract (ISM-1451). Where a provider will not commit, send it no organisational data unless the data owners consent.

  3. Get informed, explicit consent before using data to train or improve a model.

    Ask the data owner for each data set, which for customer data means the customer. Tell them the data, model, purpose and who will hold the resulting model, and record their clear yes and its date. A default opt-in or general terms are not enough.

  4. Gate your own training pipelines on recorded consent.

    Before a data set drawn from AI application data is used to train, fine-tune or improve a model, check that consent covering that data, model and purpose exists and predates the use, and exclude anything without it.

Keep it running

Recheck each AI service's data-use settings and terms when the vendor updates them, adds features or you change service tier, because defaults can change. Consent covers only the data, models and purpose it names, so a new purpose or model needs new consent.

Evidence and audit

  1. Check 1 of 3What an auditor asks for

    Ask for

    The model APIs and AI services used by each AI application the organisation develops, with their data-use settings, compared with each application's architecture and the subscription records for AI services.

    Look at

    Whether every model API and AI service is listed, and whether its live setting or contract stops organisational data being used to train, fine-tune or improve models.

    Good evidence

    Every model API and AI service used by an AI application the organisation develops is listed, and each one's live setting or contract excludes training, fine-tuning and model improvement on organisational data, except where recorded consent covers that use.

  2. Check 2 of 3What an auditor asks for

    Ask for

    The data sets used in recent training, fine-tuning or model improvement work, traced to their sources.

    Look at

    Whether any came from AI applications, and whether each such source has consent dated before the use.

    Good evidence

    Every data set drawn from AI application data has consent from its data owners dated before it was used.

  3. Check 3 of 3What an auditor asks for

    Ask for

    A sample of consent records, with what the data owners were told.

    Look at

    Whether consent came from the actual data owner as an active choice, and whether they were told the data, model and purpose.

    Good evidence

    Each sampled consent was given by the data owner as a clear yes, not a default or general terms, and names the data, model and purpose it covers.

Mappings

Closest match in ISO/IEC 27001: A.5.34, Privacy and protection of personal identifiable information (PII), which partially overlaps with this control.

ISO 27001

ISO/IEC 27001: 1 mapping

Mappings show how controls relate across frameworks. They do not imply full equivalence or certification.