Defence AI assurance starts with named responsibility

The UK’s JSP 936 directive turns broad AI principles into a lifecycle governance task for suppliers and public bodies.

An AI supplier can describe a reliable model while leaving the buyer uncertain about who owns the resulting service. Defence assurance has to close that gap. A public body needs a clear purpose, named responsibility and evidence that remains usable when the system changes. The relevant question is whether the proposed capability can be governed throughout its life, including ordinary maintenance and eventual retirement.

The United Kingdom published [JSP 936 Part 1](https://www.gov.uk/government/publications/jsp-936-dependable-artificial-intelligence-ai-in-defence-part-1-directive) on 13 November 2024. The directive addresses governance, development and assurance across the AI lifecycle. It builds on the [June 2022 defence AI strategy](https://www.gov.uk/government/publications/defence-artificial-intelligence-strategy/defence-artificial-intelligence-strategy) and the [associated responsible-use policy](https://www.gov.uk/government/publications/ambitious-safe-responsible-our-approach-to-the-delivery-of-ai-enabled-capability-in-defence/ambitious-safe-responsible-our-approach-to-the-delivery-of-ai-enabled-capability-in-defence). These dated public documents establish a policy approach. They do not certify any supplier, model or programme as compliant.

A safe public example is an original illustrative assistant that helps procurement staff find clauses in maintenance contracts. It can suggest relevant passages and identify the document version. It should not approve a purchase or reinterpret an agreement on behalf of an officer. The scope is deliberately narrow because errors can be checked against the original records before a consequential decision follows.

The buyer can ask the supplier to show what happens when the source is missing, a document has been superseded or two clauses appear inconsistent. A useful output reports the uncertainty and links to the material. A convincing but unsupported answer should fail acceptance. This proposed test concerns administrative document handling; it does not describe an operational military system.

Named responsibility should extend beyond the project sponsor. Someone owns the approved document collection. Someone accepts model changes. Someone can suspend the assistant when outputs become unreliable. Those roles need time and authority, plus a shared understanding of when to escalate. Writing a name in a register does little if the named person cannot obtain evidence or change the service.

Assurance evidence needs a defined context. A supplier’s general benchmark may say little about a contract archive containing old scans, unusual abbreviations or restrictive handling requirements. In the illustrative pilot, reviewers can prepare a controlled sample reflecting those conditions and withhold some items from development. Record the accepted result, the source reference and the reason for every correction.

The contract should address how evidence transfers. A new supplier may need the permitted document set, model configuration, test cases and unresolved issues. A handover that contains only a dashboard and summary accuracy figure will make continuity difficult. Intellectual-property rights, licences and data-handling restrictions need explicit review. The public policy encourages accountable adoption; it does not eliminate those purchasing questions.

Human review also needs evaluation. Staff may accept an answer because it is clear and fast, even when the source contradicts it. The pilot can include plausible errors and check whether reviewers detect them. If they do not, the design may need stronger source presentation or a smaller task. A nominal approval step is weak evidence of effective oversight.

Routine change can undermine an otherwise careful acceptance. A supplier update, new archive format or different user group may alter the risk. Keep a change record and specify which changes require renewed review. Retain the ability to return to a known configuration or stop the affected feature. These measures preserve accountability without assuming that every change has the same significance.

The potential benefit is quicker access to checked administrative information and fewer repeated searches. The failure is a service whose confidence outpaces its evidence, or whose owner cannot intervene. JSP 936 provides a public reference for treating that issue as a lifecycle responsibility. Its value should be assessed through the buyer’s actual governance and tests, with policy documents, supplier claims and observed acceptance kept distinct.

A practical review meeting can request one difficult case rather than another slide deck. Trace the request, the retrieved material, the model’s response and the officer’s correction. That single chain can reveal missing version information or an unclear escalation path before routine use makes the gap expensive.