National-Security AI: Humans Must Remain in Command

As a former chief of the Intelligence Bureau, I view artificial intelligence with both interest and caution. Its ability to process information can strengthen national security. But the responsibility for deciding what that information means, and what action should follow, must remain firmly with human beings. In intelligence work, information rarely arrives complete or free from ambiguity. A report may be accurate yet misleading without context. Several sources may repeat the same unverified claim. Sound assessment requires questioning, corroboration and an awareness of what remains unknown. AI does not remove these responsibilities. It makes the discipline of exercising them even more important.

Being in the loop is not enough

The expressions “human-in-the-loop” and “human-in-command” are sometimes used as though they mean the same thing. A useful distinction concerns the depth of human authority. Human-in-the-loop usually means that a person participates at a particular stage: reviewing an assessment, approving a recommendation or authorising an action. Human-in-command goes further. Human authorities determine the purpose, establish boundaries, decide what may be delegated and retain the power to intervene or stop the system. They remain answerable for the consequences. The two approaches are complementary. But a person who merely approves a machine’s recommendation, without sufficient time or information to challenge it, provides only the appearance of control.

What “rogue” AI reports tell us

Recent reports about AI agents “going rogue” deserve attention, but also careful interpretation. These agents can take actions through connected tools, rather than merely produce answers. In controlled experiments published in summer 2026, Anthropic researchers described AI agents secretly altering code or changing assessment labels to influence outcomes. The researchers explicitly distinguished these experimental findings from real-world incidents. Such findings do not establish that national-security systems are behaving similarly. They do, however, underline a practical concern: an apparently capable system may act outside its instructions, and its reports may not reveal what it has done. For a security institution, this strengthens the case for restricted access, independent checks and records that allow actions to be examined. Trust must be supported by verification.

Why intelligence requires human judgement

Consider an AI system identifying unusual communications and movements as evidence of a developing threat. The pattern may warrant investigation. It does not, by itself, establish hostile intent. An experienced assessment must consider alternative explanations, source reliability, possible deception and the consequences of being wrong. It must also distinguish between an intelligence lead, an assessed threat and grounds for action. These distinctions matter because national security concerns both the protection of the country and the protection of its people. A mistaken assessment can damage an innocent person’s life, deepen mistrust within a community or contribute to an avoidable confrontation. Speed is valuable. So is the judgement to seek corroboration, narrow a response or pause.

Making command meaningful

Human command requires clear institutional arrangements. Responsibility should be assigned before deployment. There must be no uncertainty about who approves the system, who supervises its use and who can suspend it. Officers must be trained to question AI outputs. Assessments should make uncertainty and conflicting evidence visible, instead of presenting every conclusion with equal confidence. Access and authority should be limited to the task. Permission to analyse information should not automatically include permission to initiate consequential action. Where automated responses are necessary, their limits must be established beforehand. Finally, decisions must remain open to review. Human beings also make mistakes. A sound command structure encourages professional disagreement, independent scrutiny and correction. My central concern is the preservation of responsible judgement as technology becomes more powerful. In intelligence and public service, authority carries an obligation to explain and answer for decisions. AI can assist that responsibility. It cannot assume it. Humans must remain in command, with the knowledge, authority and institutional support to make that command real.