Defence AI: Clarity under uncertainty
AI can organise large volumes of information and reveal useful connections. In defence, its value depends on the quality of the whole decision process: what is known, what remains uncertain, and who has the authority to act on that information?
At a glance
- 01
Every material claim needs traceable evidence and clear limits on where it applies.
- 02
Human responsibility requires time, information and a real ability to intervene.
- 03
Readiness must be demonstrated for a specific purpose and its operating conditions.
Start with a bounded decision
A useful starting point is a clearly defined support task, such as reviewing maintenance records or assembling information for a procurement decision. Which details are missing today? Which errors cause delays? Which decision remains with an accountable person? These questions make the expected benefit testable and determine which data the system actually needs.
NATO’s revised AI strategy includes accountability, traceability, reliability and governability among its principles. An application design should translate those principles into a coherent relationship between purpose, users, permitted outputs and limits. A persuasive demonstration using selected documents provides only partial evidence. The proposed workflow still needs to work with the documents, constraints and people found in everyday use.
Sources: NATO: Revised Artificial Intelligence Strategy (2024)
The evidence behind an accountable decision
Source · timestamp · reliability
Context · contradiction · uncertainty
Options · authority · approval
Different evidence states remain visible.
Keep the evidence attached to the claim
A summary should show which records support its material claims, when those records were produced and how they were processed. Conflicting information deserves particular attention. Several documents repeating one original report do not provide independent confirmation. A well designed system makes that dependency visible so that repetition cannot quietly acquire the appearance of stronger evidence.
Uncertainty also needs a useful explanation. Missing data, conflicting sources and an uncertain model response are different problems with different next steps. A precise looking percentage is useful only if its meaning has been checked for the task concerned. Often the more helpful output is a clearly identified information gap, linked to the conclusion that depends on it.
Sources: NATO: Revised Artificial Intelligence Strategy (2024)NIST: AI Risk Management Framework 1.0 (2023)
Make human responsibility practical
Responsibility becomes concrete in the workflow. A reviewer needs access to the decisive evidence, a way to understand alternatives and the ability to reject or defer a recommendation. That requires enough time and the authority to request further review. Approval loses its value when an interface presents only a finished answer and makes disagreement unnecessarily difficult.
For an illustrative document review system, the division of responsibility could be straightforward: AI flags inconsistencies and prepares a reasoned draft. A domain specialist assesses the facts, while changes to authoritative records follow the established approval process. The record connects the information used, the system version, the proposal and the human decision. A later reviewer can then understand what was actually known at the time.
Sources: NATO: Revised Artificial Intelligence Strategy (2024)NIST: AI Risk Management Framework 1.0 (2023)
Test readiness with difficult cases
A credible evaluation covers routine tasks alongside incomplete, outdated and contradictory inputs. Changes in users or permissions, and the loss of a data source, also matter. Define in advance when the system should withhold a conclusion, request review or fall back to a simpler workflow. Acceptance criteria need to cover the complete system, including the people using it.
The NIST AI Risk Management Framework emphasises evaluation in the intended context and documentation of limitations. For a project, that suggests a clear chain of evidence: a defined purpose, a reproducible evaluation, a documented result and a bounded approval. When models, data or workflows change, review which findings remain valid. A responsible introduction makes the scope of the evidence and the outstanding questions equally visible.
Your next step
Define a practical evaluation framework
Discuss a bounded support task, its data requirements and the evidence needed to assess it.
Discuss your projectSources & further reading
- NATO: Revised Artificial Intelligence Strategy (2024)
Official principles for responsible AI, data quality and evaluation. The applications in this article are illustrative design proposals.
- NIST: AI Risk Management Framework 1.0 (2023)
A general framework covering context, reliability, human responsibility and evaluation. It does not certify or approve a particular system.