In Parts 1 and 2 of this series on AWS, we examined whether existing international humanitarian law (IHL) adequately governs autonomous weapon systems (AWS) and addressed the accountability gap that critics invoke to demand a new international treaty. The answer to both questions was essentially the same: current law is sufficient, and the alleged gaps are smaller than the disarmament lobby suggests.
Part 2 concluded by identifying a more operationally relevant question: what level of human involvement actually ensures the lawful and effective use of AWS? This Part 3 takes up that question.
Estimated reading time: 13 minutes
I. The Wrong Debate Is Costing Us the Right Answers
Disarmament NGOs largely drive the dominant narrative. This narrative frames the issue as a choice between ‘killer robots’ without human oversight. It contrasts this with systems in which a human approves each engagement in real time. Both ends of this spectrum are unhelpful. The first is a strawman. The second would render AWS operationally useless in precisely the scenarios where they are most needed.
The more productive question — the one that actually serves defence decision-makers — is how actors structure human control, at which stages of a system’s lifecycle they exercise it, and who exercises it. This is, in turn, what this post addresses. This reframing builds directly on Part 2’s conclusion that accountability already exists across multiple legal and organisational layers. The question is how actors operationalise that accountability in system design and use.
II. A Term That Keeps Evolving – and What That Tells Us
From MHC to “Context-Appropriate Human Judgement and Control”
The concept of “meaningful human control” (MHC) has dominated academic and diplomatic discussion of AWS regulation for over a decade. However, the term itself has recently shifted. In the United Nations Group of Governmental Experts (GGE) on lethal autonomous weapon systems (LAWS), the May 2025 rolling text replaces “meaningful human control” with the phrase “context-appropriate human judgement and control.”
This is not merely terminological housekeeping. The shift reflects a substantive disagreement that has paralyzed GGE negotiations for years. Some states read ‘meaningful human control’ as implying a fixed, universal standard. In practice, this would require real-time human involvement in every targeting decision.
Consequently, that reading would make many legitimate AWS configurations legally suspect and was therefore unacceptable to states with advanced autonomous weapons programmes. The new formulation builds the word ‘context’ directly into the standard. It acknowledges that what counts as adequate human control depends on the weapon system, the operational environment, the type of target, and the phase of the engagement cycle.
Towards Human Judgement And Control Across The AWS Lifecycle
The GGE’s May 2025 rolling text gives concrete expression to this context-sensitivity. It requires states to ensure that LAWS effects are “adequately predictable, reliable, traceable and explainable to those responsible for their use.” In addition, it mandates rigorous testing and evaluation. This includes “realistic simulations of its use in operational environments, to enable a human user to have a reliable expectation of how LAWS will perform in the anticipated circumstances of its use” (GGE Rolling Text, May 2025, Sections III. 6. A and IV. 5).
Taken together, these lifecycle obligations confirm that the design, testing, and acquisition process provides for human judgement and control not only at the moment of engagement, but also – and primarily – before it.
Is “context-appropriate human judgement and control” a better standard than “meaningful human control”? For defence practitioners, yes. First, it more honestly reflects the operational reality that this formulation distributes control across a system’s lifecycle, rather than concentrating it at the moment of firing. It also maps directly onto existing state practice. Moreover, the 2023 update to U.S. DoD Directive 3000.09 — the foundational U.S. policy on autonomy in weapon systems — requires designers to ensure that AWS allow commanders and operators to exercise ‘appropriate levels of human judgment over the use of force’ (DoD Directive 3000.09, January 2023, Section 1.2.a). Here, the word “appropriate” is doing important work: it signals a flexible, context-sensitive standard rather than a fixed universal requirement, consistent with the direction the GGE’s own language has taken.
The GGE’s new language and the DoD’s existing policy have quietly converged on the same concept. That convergence matters for defence contractors and programme managers. The regulatory direction of travel is context-sensitivity and lifecycle governance, not a blanket real-time control requirement.
III. Why “Human in the Loop” Is the Wrong Standard
The ‘human in the loop’ (HITL) model is superficially appealing. It requires a human operator to approve each individual engagement in real time. It appears to guarantee human oversight and preserve individual accountability. We should take its concerns seriously. However, as an operational standard for AWS, it fails on multiple levels.
In Combat, Speed Is Of The Essence
First, it is a speed problem. In practice, designers develop many AWS precisely for threat environments where engagement windows last only seconds or fractions thereof. These include terminal air defence, counter-drone operations, and loitering munitions responding to rapidly emerging targets. A mandatory confirmation loop does not add meaningful oversight in these scenarios; it eliminates the capability.
The HITL Model adds Cognitive Workload Instead Of Reducing It
Second, it creates a workload trap. Requiring real-time human approval under high-tempo, degraded communications and information-saturated conditions shifts cognitive load back to the human operator at the worst possible moment. Battlefield conditions undermine precisely the decision quality that HITL aims to guarantee.
The HITL Model Distorts Human Accountability
Third — and most importantly — it distorts accountability rather than preserving it. The NATO Science and Technology Organization’s December 2025 report on Human Systems Integration for Meaningful Human Control (TR-HFM-330) makes this point directly:
“Making people accountable without giving them the ability, autonomy, and transparency, i.e., using them as a moral crumple zone would be extremely unfair and should be avoided” (NATO TR-HFM-330, 2025, PDF p. 87).
Human-robot interaction research introduced the term ‘moral crumple zone’ to describe a human nominally responsible for a decision it could not realistically influence. This is not accountability. It is a liability transfer mechanism that creates legal exposure without adding genuine oversight. This concern directly echoes Part 2’s discussion of “blessed ignorance” and the need to ensure that this framework pairs responsibility with real informational and decision-making capacity.
Human Accountability Must Remain Genuine
The 2018 UK working paper to the GGE reached a closely related conclusion. The UK argued that
“the level, nature and primacy of human control over specific functions is the key consideration in the LAWS debate rather than technology” (UK GGE Working Paper, CCW/GGE.2/2018/WP.1, Section III. para. 8). It also suggests that “taking a human-centric, through-life approach enables human control to be considered at multiple stages and from various perspectives” (ibid., IV. para. 12).
The paper’s detailed lifecycle framework – covering six stages from political control through post-deployment analysis – shows that distributing human control across the weapon system lifecycle does not dilute responsibility. Instead, actors build accountability into each stage. In particular, this includes requirements definition and system design, as well as legal reviews, training, rules of engagement, operational deployment, and in-service feedback.
The question, therefore, is not whether humans must be present at the moment of engagement, but how actors structure human control across the full lifecycle of the system. The aim is to ensure that accountability remains genuine rather than nominal.
IV. MHC as a Systems Engineering and Procurement Discipline
This is where the NATO TR-HFM-330 report makes its most important practical contribution. The report was published in December 2025 by a multinational research group. Contributors included the U.S. Air Force Research Laboratory, TNO Defence (Netherlands), Fraunhofer (Germany), Cranfield University (UK), and others. It approaches MHC not as a political concession but as a systems engineering challenge.
The Core Findings of the NATO Report
Its central finding is direct: MHC is not a binary variable — it does not exist as either “present” or “absent.” It does not manifest at a specific moment or place. Instead, it emerges through a series of interactions over time across the full lifecycle of the weapon system. As the report concludes, this approach “sets itself apart from other approaches to ensuring MHC, such as relying on high-level platitudes, technological panaceas or simply prohibiting all “killer robots.” The report treats MHC as a complex, multifaceted challenge. It requires further exploration to identify clear design and operational guidelines.
The report identified four nested control loops. Each represents a distinct form of human oversight exercised by different actors at different timescales. In the report’s own words, these comprise:
- “The operator loop, wherein an operator governs the AI system through a control unit.”
- “The mission loop, overseen by mission commanders who determine the deployment strategy and locations for the AI system.”
- “The design loop, managed by AI engineers responsible for designing the system, its interfaces, and training the machine learning models.”
- “The governance loop, led by policy makers who develop laws, protocols, and rules of engagement that impact the AI system.”
(NATO TR-HFM-330, 2025, SOTEF model section, PDF p. 41).
Methods for Human Control Over the System Lifecycle
The NATO report also identified 17 “candidate methods” to ensure meaningful human control over AI-based systems. These methods operate in design, acquisition requirements writing, training, and mission planning. As the report’s Executive Summary states, ” These potential solutions vary from design guidelines to situational awareness metrics and organizational training” (NATO TR-HFM-330, 2025, Executive Summary, p. 15).
For defence industry decision-makers, this lifecycle framing translates directly into procurement and engineering practice. By way of example, the NATO report provides detailed guidance on the following methods to ensure MHC over AWS:
1. MHC requirements for acquisition contracts
The report provides validated ‘shall’ and ‘should’ requirement templates for system acquisition contracts from the outset. These templates cover situation awareness provision, system explainability, operator intervention capability, temporal and geographic constraints, and after-action traceability. These are not aspirational standards; they are testable, measurable criteria (NATO report, PDF pp. 57, 58).
2. Advance Control Directives
Commanders define the acceptable behavioral envelope of an AWS during training and pre-deployment exercises. In doing so, they effectively program their command intent into the system before operations begin. This preserves genuine human control without requiring real-time intervention in execution (NATO report. PDF p. 60).
3. Responsible Neglect Tolerance (RNT)
This metric was developed within the HFM-330 group. It defines how long a system can operate without human oversight or intervention while remaining within ethical and legal bounds. It also indicates when human intervention becomes necessary. This is a direct design input for engineers. It shows how much autonomy a system can responsibly exercise in a given operational context (NATO report, PDF p. 64).
This short list is not conclusive. The NATO toolkit provides many more methods to implement human control over AWS throughout the system lifecycle. These methods can be seen at a glance in Section 4.3 of the report (PDF pp. 54, 55) and explored in more detail in the candidate method descriptions in Section 4.4. (PDF pp. 55 et seq.).
V. A Case Study on Dynamic Targeting And “Distributed Human Control”
Reframing the Question Of Meaningful Human Control
This lifecycle approach finds important support in operational scholarship on military practice. Merel Ekelhof conducted research on human control in targeting, commissioned by the Netherlands Ministries of Defence and Foreign Affairs. She argues that the standard MHC debate asks the wrong question. Rather than asking whether an individual operator exercises meaningful human control over a weapon at the moment of engagement, Ekelhof contends:
“we should ask how the military organization as such can or cannot ensure meaningful human control over important targeting decisions” (Ekelhof, ‘Moving Beyond Semantics on Autonomous Weapons: Meaningful Human Control in Operation’, Global Policy, Vol. 10, No. 3, 2019, p. 347).
This reframing maps directly onto the NATO report’s lifecycle architecture and directly challenges the operator-centric HITL model.
Human Control Is Excercised Even Before a Single Shot Is Fired
Ekelhof’s analysis of conventional air operations illustrates why. Drawing on interviews with over 90 practitioners and participant observation of targeting exercises, she demonstrates that
“in practice, various human beings exercise different forms of control at various junctures in the decision-making cycle that, eventually, leads to the deployment of weapon (systems).”
The consequence is that “the human role in formulating objectives, developing and approving targets, conducting collateral damage estimations, weaponeering solutions and proportionality analyses, and deciding on operational constraints and restrictions is ignored” when MHC is defined solely by the operator-weapon relationship at the moment of firing (ibid., p. 347). In other words,
“human control does not need to have a direct link with the weapon system” — humans “typically exercise different forms of control over important decisions (such as target selection and engagement), even before weapons are activated” (ibid., p. 347).
This directly reinforces the NATO report’s finding that the vast majority of the 17 candidate MHC methods it identifies operate before the weapon is fired (“left f launch”), and confirms that the governance and design loops are not peripheral to MHC but central to it.
VI. Four Practical Implications for Decision-Makers
The practical implications below draw on the NATO TR-HFM-330 report. The report uses the established term ‘MHC’ throughout, reflecting its three-year research timeline.
The substance maps directly onto what the GGE’s May 2025 rolling text now calls ‘context-appropriate human judgement and control.’ The procurement advice that follows applies equally under either formulation.
The NATO framework generates four concrete implications for defence programme managers, engineers, and legal advisers.
1. Build MHC into Contracts, Not into Apologies.
MHC requirements belong in system specifications and acquisition contracts from day one. Retrofitting them after development is costly, unreliable, and legally precarious. The NATO HFM-330 requirement templates and the GGE’s lifecycle measures (Section IV of the May 2025 rolling text) provide a ready-made checklist.
2. Treat ROE and Mission Parameters as Forms of Human Control
The GGE rolling text explicitly requires that mission parameters regarding target types, duration, geographic scope, and operational scale “cannot be modified by the system without context-appropriate human judgement and control” (GGE Rolling Text, May 2025, Section III. 6. C). Designing these constraints in — and documenting that design — is a legal review obligation under Article 36 AP I. It is also a command accountability measure. This aligns with our Part 2’s conclusion that flexible instruments such as ROE and FRAGOs are central to governing AWS in practice.
3. Distinguish Genuine Accountability from Nominal Sign-Off
Some Systems require human confirmation but give operators no real time, information, or intervention capability to act on that confirmation. These systems do not preserve accountability. Instead, they create it on paper while transferring risk to the individual operator. This is the moral crumple zone problem. Legal advisers should flag it in acquisition reviews; engineers should design against it.
4. Invest in Operator Training as an MHC Measure
DoD Directive 3000.09 requires that operators and commanders “understand the functioning, capabilities, and limitations of a system’s autonomy under realistic operational conditions” through training, TTPs, and doctrine that are established and periodically reviewed (DoD Directive 3000.09, January 2023, Section 2.9. b. 8). The NATO report goes further. It envisions “extensive training of both the human operators and the actual AI systems, together, in highly realistic simulated situations” (“human-machine co-training”, NATO report PDF p. 68).
VII. Conclusion
The MHC debate has been framed for too long as a tension between legal compliance and operational effectiveness. The evidence surveyed across this series does not support that framing.
As Parts 1 and 2 have shown, existing IHL already governs AWS and provides a sufficient legal framework for their use. The GGE’s evolving language confirms that the relevant standard is context-sensitive, lifecycle-spanning, and consistent with the kind of structured autonomy that modern defence systems require. NATO’s own research demonstrates that MHC is achievable through engineering discipline, procurement standards, and training. It is not achieved through the operational fiction of a human approving every engagement in real time.
The states and organisations that will navigate the regulatory environment most successfully treat MHC as a systems engineering and governance discipline from programme inception. They must also demonstrate, through documented design choices, testing records, ROE, and training programmes, that their systems operate within a genuine chain of human command and control.
Legal compliance and defence readiness are not opposites. In properly designed and employed AWS, they converge. The debate, therefore, should move beyond abstract calls for prohibition and focus instead on how to implement context-appropriate human judgement and control in practice.