Defense AI: Application Areas and Principles of Human Control

What defense AI is and why it is needed

A UK Ministry of Defence policy paper source defines AI as 'a family of general-purpose technologies'. These are technologies that let machines do tasks that would otherwise require human or biological intelligence, especially technologies that learn how to do so from data. The paper states that AI will improve decision-making and the conduct of operations by letting people handle volumes of data that were previously hard to imagine. At the same time, it is firm that AI is not an end in itself. Its position is that the case for developing and using AI is strong when there is a clear benefit or when it leads to more ethical outcomes.

NATO's summary of its revised AI strategy, released on July 10, 2024 source, observes that powerful technologies such as generative AI have spread at an unprecedented pace. It therefore states that AI should be adopted as quickly as possible wherever it can be used. At the same time, the strategy includes identifying and countering threats from adversaries' use of AI among its goals. In other words, defense AI has to address two directions at once: 'the AI we use' and 'the AI our adversaries use'.

What most sets defense apart from other industries is that decisions lead directly to the lives of soldiers and civilians. The UK Ministry of Defence source acknowledges that the risks and concerns about AI's impact on people may be especially high in defense. It warns that failing to address them could cost public consent and even weaken operational capability. That is why, in discussions of defense AI, 'who decides and who is responsible' matters as much as performance.

Components: application areas and levels of autonomy

The areas where defense AI is used are usually divided into four. First, intelligence, surveillance and reconnaissance (ISR) analysis organizes accumulating material such as imagery, signals and reports, and shortlists candidates that analysts should look at first. Second, maintenance and logistics are relatively low-risk support tasks, such as anomaly detection to spot signs of equipment faults or time-series forecasting to predict demand for parts. Third, command decision support summarizes the situation and presents options with their rationale, while the commander makes the final decision. Fourth, unmanned systems take forms such as unmanned aerial vehicles (UAVs), unmanned underwater vehicles (UUVs) and unmanned ground vehicles (UGVs), as the Wikipedia article source describes. The sources this note is based on contain no performance figures by area. Individual effects should therefore be judged from each program's test results.

NATO source distinguishes narrow AI built for specific tasks from frontier models and foundation models used for many purposes. The two differ in risk, the actors involved, and level of complexity. NATO also highlights as key issues the technical and governance problems of applying dual-use (civilian and military) technologies from the commercial market in military settings, and questions of responsibility when humans and machines work as a team.

Autonomy is not a matter of 'present or absent' but of degree. The UK Ministry of Defence source explains that the human involvement required for a weapon may be real-time supervision, or it may take the form of people setting the system's operational parameters (when, where and against what it operates). According to the UK Ministry of Defence distinction cited by Wikipedia source, automated systems perform complex tasks but cannot make complex decisions. Autonomous systems, by contrast, can decide on a course of action without human supervision. Wikipedia source notes that, as of 2025, most military drones and robots still require human control and are not truly autonomous systems.

Design principles: responsible AI at NATO and the UK Ministry of Defence

In its 2021 AI strategy, NATO adopted six Principles of Responsible Use (PRU) for AI in defense source: lawfulness, responsibility and accountability, explainability and traceability, reliability, governability, and bias mitigation. The revised strategy goes beyond declaring the principles and aims at 'operationalizing the PRU', turning them into actual procedures. Its expected outcomes include more AI use cases that comply with the principles, and measurable steps in the NATO Defence Planning Process for integrating AI into Allied forces on the basis of high-quality data.

The UK Ministry of Defence source sets out three pillars: 'ambitious, safe, responsible'. Central to this is a 'systems approach'. Because the problem may lie less in 'what' a capability is designed to do than in 'how' it does it, the ministry seeks solutions across the whole lifecycle, from design and manufacture to operation and disposal, and across the 'system of systems' in which many systems interact. The paper states that, since the technology changes quickly, it will produce safe and responsible outcomes through defense-wide processes rather than setting arbitrary restrictions in advance. It also commits to always meeting domestic and international legal obligations, continuing to apply existing safety and regulatory frameworks, and having clear frameworks and processes for ethical adoption. The excerpt this note draws on does not include the UK's detailed list of principles, so those details should be checked in the original document.

The two organizations phrase things differently, but the common ground is clear: comply with the law, make clear who is responsible for outcomes, be able to trace what happened and why, operate reliably, and let people control and stop the system. In practice, the starting point is rewriting these principles as requirements and test items.

The international debate on lethal autonomous weapon systems (LAWS)

The first reason the debate is hard is that definitions differ. According to the Wikipedia article source, there is no generally agreed definition of LAWS among countries. US Department of Defense Directive 3000.09 defines it as 'a weapon system that, once activated, can select and engage targets without further intervention by a human operator', and this includes systems that operators can supervise and intervene in. France takes a narrow view, limiting it to 'fully autonomous' systems that can change their own operating rules beyond a defined framework of use. Still, many countries and organizations share a core feature: 'once activated, it selects and engages targets without human input at that moment'. According to the same article, the UN Group of Governmental Experts (GGE) made progress at its September 2026 meeting by agreeing on wording that characterizes LAWS.

The International Committee of the Red Cross's 2021 position paper (ICRC, w3) describes how autonomous weapons work as follows. After a person first activates it, the system initiates attacks on its own based on sensor information and a generalized 'target profile'. As a result, the user neither chooses nor knows the specific target or the precise time and place. The ICRC considers that this risks harming both civilians and combatants and escalating conflict. It also points out that it makes compliance with international humanitarian law harder and raises the ethical problem of replacing human decisions about life and death with sensors and software. The ICRC therefore recommended new legally binding rules. Autonomous weapons whose effects cannot be sufficiently understood, predicted and explained, and autonomous weapons that target people, should be prohibited. The rest should be regulated through requirements on target types, duration, geographic scope and scale of use, situations of use (such as places without civilians), and human-machine interaction (effective supervision and timely intervention and deactivation).

The UK Ministry of Defence source does not rule out putting AI into weapon systems. However, it acknowledges that some concepts may be impossible to implement safely and responsibly. It also states that weapons that identify, select and attack targets must have 'context-appropriate human involvement'. The discussion often called 'meaningful human control' ultimately deals with this point: what people know and decide, when, and to what extent. Meanwhile, Wikipedia source notes that in some defensive systems that must respond quickly to incoming threats, that speed has made it difficult to keep humans in the decision loop. The fact that the form of control can vary by situation should also be kept in view.

Step-by-step application: testing, evaluation and the adoption process

First, define the purpose. As the UK Ministry of Defence source says, AI is not an end in itself. Start by writing down which bottleneck in which task it reduces, and whether it is clearly better than doing the work without AI. Second, identify safety, legal and ethical risks early. The UK Ministry of Defence source states that it will provide frameworks to help development teams find and resolve such risks early. It is realistic to set risk levels and review intensity differently depending on the application area (whether it is maintenance support or a target-related function).

Third, manage the data. NATO source treats high-quality data as a precondition for AI integration and flags as an issue that the global supply of high-quality public data for model training may shrink. Record data sources and data lineage, and check for possible manipulation such as data poisoning. Fourth, test and evaluate rigorously. The UK Ministry of Defence source states that it will remove problems at the design stage (design out) and rigorously test AI-based solutions. Include in the test items the gap between lab performance and real-world conditions, situations where an adversary deliberately tries to deceive the system such as adversarial examples, and red-team attacks.

Fifth, design human control. As in the requirements set out by the ICRC source, confirm that effective supervision and timely intervention and deactivation are actually possible. Document whether the rationale is visible on screen, whether the stop mechanism works, and who gives approval. Sixth, keep monitoring after deployment. NATO source has set the goal of protecting and monitoring AI technologies and managing the associated risks. Keep re-evaluating to check that performance is not degrading due to data drift.

General frameworks from the civilian sector are also useful references. According to the US NIST page source, the NIST AI Risk Management Framework (AI RMF) was released on January 26, 2023. It is designed for voluntary use and helps build trustworthiness considerations into the design, development, use and evaluation of AI. NIST released a generative AI profile on July 26, 2024, and a concept note for a critical infrastructure profile on April 7, 2026. AI RMF 1.0 is currently being revised. Because this page does not address defense specifically, it should be used as a reference framework for designing risk management processes, not as a substitute for defense-specific legal and rules-of-engagement review.

Checklist

Before adoption, confirm the purpose and responsibility. It should be documented whether the system provides a clear benefit and who is responsible for its outcomes. Also check whether NATO's six principles source have been translated into requirements and whether a review of domestic and international legal obligations source has been carried out.

On the technical side, confirm that the sources and quality of training and evaluation data can be traced. Also check whether the system has been tested under realistic and adversarial conditions, and whether it is possible to explain why a result came out as it did and to trace the records.

On the control side, confirm that people set the operational scope (targets, time, area) and that real-time supervision, intervention and stopping are possible sourcesource. It is also important that operators have the time and information to critically review the AI's suggestions. After deployment, check that there are processes for monitoring performance changes, reporting incidents and re-evaluation.

Common mistakes and limitations

The most common mistake is to treat autonomy as the same thing as capability. The UK Ministry of Defence source states that 'more autonomous' does not necessarily mean 'more capable', and that Human-Machine Teaming, in which people and machines work together, produces the best results in terms of overall effectiveness, use of resources and realism. The second mistake is inflating test-environment performance as if it were combat performance. The battlefield differs from the training data, and adversaries deliberately try to deceive. The third is when people approve only as a formality. Once a habit of approving the machine's suggestions as-is sets in, actual control weakens even if, on paper, a person made the decision.

There are also structural limitations. As Wikipedia source shows, definitions of 'autonomy' and 'LAWS' differ by country, so people end up discussing different things under the same word. As the ICRC source points out, with systems that attack on their own according to a target profile, the user does not know the specific target or timing. That raises the risk that responsibility becomes blurred. The issues NATO source identifies also remain unresolved: governance problems when using civilian dual-use technologies in the military, the energy consumption of compute-intensive AI, and the shrinking supply of high-quality public data.

Finally, there are limits to the available information. Much of defense AI is not public, so it is hard to verify actual performance and levels of control from the outside. This note is also based on published policy documents and position papers, so it is not sufficient grounds for judging how any specific system actually operates.

Further study

Start with the UK Ministry of Defence's 'ambitious, safe, responsible' policy paper source to get the big picture of how a defense organization tries to handle AI adoption and safety together. Then use the summary of NATO's revised AI strategy source to review principles and goals at the alliance level.

For the autonomous weapons debate, the ICRC position paper source presents the issues and regulatory proposals most concisely. Skim the Wikipedia LAWS article source for differences in national definitions and recent UN discussions, but since it is a secondary source, it is best to check the cited originals (the US Department of Defense directive, GGE reports) separately.

If you need to actually design a risk management process, go from the NIST AI RMF page source to the framework text, the playbook and the generative AI profile. Related terms worth reading are human-in-the-loop, AI risk management, red teaming and adversarial examples.