NATO’s AI interoperability problem includes governance

The 2024 strategy puts shared testing, data quality and responsible-use principles alongside technical integration.

Two organisations can exchange an AI output successfully and still misunderstand each other. A category may have a different meaning, a confidence score may come from a different method or a data field may describe a different period. Interoperability therefore includes more than compatible software. It requires enough shared understanding to decide whether another organisation’s result can be used responsibly.

NATO’s [revised AI strategy summary](https://www.nato.int/en/about-us/official-texts-and-resources/official-texts/2024/07/10/summary-of-natos-revised-artificial-intelligence-ai-strategy), released on 10 July 2024, connects AI adoption with interoperability, data quality, workforce readiness and testing. The [announcement](https://www.nato.int/en/news-and-events/articles/news/2024/07/10/nato-releases-revised-ai-strategy) establishes that policy update. The [Data and Artificial Intelligence Review Board’s remit](https://www.nato.int/en/about-us/official-texts-and-resources/official-texts/2022/10/13/natos-data-and-artificial-intelligence-review-board) describes a governance forum and responsible-use work. These public statements establish priorities, not proof that all relevant systems already interoperate.

An original illustrative example uses a non-operational equipment-maintenance catalogue. One office describes a component as needing review; another uses the same term for a confirmed fault. An AI assistant combines the records and produces a neat summary. The risk comes from collapsing two different meanings into one label. A shared format prevents a parsing error but does not prevent that semantic mistake.

A proposed pilot can begin with a common data dictionary. Define the meaning, unit, date and source of each field that crosses the interface. Record whether an entry is observed, inferred or manually confirmed. The receiving organisation needs to know which values it can treat as authoritative and which require review. This is a general administrative illustration, not an account of NATO’s internal implementation.

Testing should involve the full exchange. Prepare a controlled case in one system, transfer it through the approved interface and ask a reviewer in the receiving system to explain the result. Include missing dates, incompatible units and ambiguous status labels. A successful test demonstrates that people and software preserve meaning together. It does not merely show that a connection returned a response.

The model adds another layer of uncertainty. A generated summary may omit qualifiers that were present in the structured record. Require it to preserve distinctions that matter to the task, such as proposed versus approved. Reviewers can inspect source references and see unresolved conflicts. If the assistant cannot explain which source supports a sentence, it should not convert the sentence into an accepted record.

Governance differences also matter. One organisation may permit a model to draft a recommendation while another requires manual review before the recommendation is shared. The interface should make those states explicit. A recipient should not assume that an incoming output has been approved because it arrived through an authenticated channel. Identity establishes who sent it; approval establishes which process it passed.

The strategy’s attention to shared testing and responsible-use work offers a public basis for this discussion. A common assessment vocabulary can help buyers compare evidence and identify missing checks. It should not erase national responsibilities or create an unsupported claim of universal certification. A process standard and a measured outcome have different roles.

Supplier changes need coordinated handling. If a model update changes summary wording or the meaning of a confidence value, recipients may need to retest the integration. Keep version information with the exchange and notify the owners of affected workflows. An isolated upgrade can cause a distant organisation to make a different interpretation without realising that the source changed.

The potential benefit is reliable sharing of information and less repeated manual reconciliation. The failure is a fast exchange that spreads ambiguity at scale. NATO’s 2024 strategy treats interoperability as part of responsible adoption. In a practical administrative service, that means agreeing on meanings, tracing sources and testing the receiving experience, while keeping the public policy separate from claims about classified or operational performance.

An acceptance record can include the receiving reviewer’s explanation in ordinary language. If that explanation differs from what the sender intended, the integration still has work to do. This gives technical teams a concrete defect to fix and gives governance teams evidence that the shared information remains usable.