The outcome, not the mechanism

A control objective says what should be true. A control says how. "Unauthorised access to model weights is prevented" is an objective; "role-based access control on the artefact store, reviewed quarterly" is a control.

Keeping them separate matters for a practical reason: the mechanism changes and the objective usually does not. When a control is replaced, the objective is what establishes that the replacement is adequate. Programmes that document only controls have to re-derive the reasoning each time something changes, and generally do not.

Objectives are conventionally grouped by what they do about a risk — preventive (stop it occurring), detective (notice it has), corrective (restore afterwards). Most risks need at least two, and a register where every objective is preventive is describing an intention rather than a control environment.

What makes one testable

An objective is only worth writing if you could tell whether it was met. Three properties do the work:

  • A named risk. An objective floating free of the risk it addresses cannot be prioritised or retired.
  • An observable condition. Something that is either true or false of the system, not a direction of travel.
  • A source of evidence. Name where the proof would come from before writing the objective, not when an auditor asks.

The commonest failure is an objective that is really a value: "AI systems are used responsibly." Nothing can falsify it, so nothing ever does.

Why AI makes this harder

Conventional control objectives assume deterministic behaviour: the control either operated or it did not. AI systems break that assumption in three ways, and each needs a different response.

Behaviour is probabilistic. "The model does not produce discriminatory outcomes" is not a testable objective, because a single counter-example neither proves nor disproves it. The testable form specifies a threshold, a metric, a population and a measurement frequency — which makes the objective inseparable from the organisation's risk appetite.

The system changes without a change. An objective met at deployment can be unmet three months later with no action by anyone, because the model drifted or the provider updated it. Objectives for AI therefore need a re-verification cadence attached, not just a verification.

Evidence is generated, not filed. For a conventional control the evidence is usually a record someone kept. For an AI control it is frequently telemetry that must have been captured at the time. An objective whose evidence source is "we would look at the logs" fails the moment the logs turn out not to contain the assembled context. See monitoring and logging.

Worked examples

Not aspirations — each names the risk, the observable condition, and where the evidence comes from.

  • Preventive. "No Tier 4 automated decision is deployed without a documented human intervention path, tested within the last six months." Evidence: the override test log.
  • Detective. "Outcome disparity between any two measured groups of ≥100 does not exceed a ratio of 1.10, assessed at each release and quarterly in production." Evidence: the fairness monitoring output. See bias testing and fairness monitoring.
  • Corrective. "Any system breaching its performance threshold in two consecutive periods reverts to human review within ten working days." Evidence: the incident record and the reversion timestamp. See incident response and rollback.

Where they come from in the frameworks

ISO/IEC 42001 does not use the phrase, but Clause 6.2's requirement for measurable AI objectives with action plans, resources, owners and timelines is the same instrument, and Annex A's controls are where the mechanisms sit against them. The Statement of Applicability is, in effect, the record of which objectives the organisation accepted and which it excluded, with reasons.

NIST AI RMF locates the equivalent in GOVERN and MANAGE, scaled to the organisation's risk tolerance rather than prescribed.

The register entry for the absence of all this is AI system lifecycle controls gap, and the ownership question underneath it is unclear AI ownership and accountability — because an objective with no owner is never evidenced.