Estimated reading time: 14 minutes
The role of the legal adviser (LEGAD) in AI-enabled operations is easy to state and hard to discharge. The adviser’s task is what it has long been: to help the commander keep planning and targeting within international humanitarian law (IHL). What has changed is the environment in which that judgement is exercised. Planning and targeting now draw on Artificial Intelligence Decision Support Systems (AI DSS), and those systems reshape when and where the adviser’s contribution matters.
The function is older than the treaty that codified it. Before 1977, armies relied on military manuals. From the 1950s, some began attaching qualified lawyers to lend informal assistance (ICRC Commentary, at para. 3341). Article 82 of Additional Protocol I (AP I) then made legal advisers a treaty requirement. That was an innovation in 1977, and it has since become settled practice (Dinstein, Legal Advisers in the Field, at p. 918). The duty now also reflects customary international law binding on States generally (CIHL Rule 141). Through each of these stages the function has held constant, and only the decision environment has moved.
Article 82 combines a firm obligation with deliberately flexible triggers. Parties “shall ensure” that advisers are available “when necessary” to advise commanders “at the appropriate level”. Dinstein reads this as a clear-cut duty tempered by unspecified conditions (Dinstein, Legal Advisers in the Field,at p. 919). Those two flexible phrases carry more weight in AI-enabled operations than they have carried before, as Section II of this post argues.
Two commitments anchor what follows. First, responsibility for decisions cannot pass to a machine; it stays with humans (GGE Rolling Text (12 May 2025), at V(1)–(2)). Second, in my view the treaty text already accommodates the new demand. The open question is institutional: whether the adviser’s involvement reaches the new substance of AI-enabled operations in practice.
I have argued that commanders are principally obliged to consider and verify AI DSS output, and I have examined where human legal judgement belongs in AI-augmented targeting. A critical literature treats the use of AI DSS primarily as a danger to be contained. I engaged that narrative in those posts and will not repeat it here. This post asks a narrower and, I think, under-examined question: what Article 82 requires of the legal adviser once operations are AI-enabled.
I. Needed more than ever — and needed early
The adviser’s involvement in targeting is not meant to begin at the moment of the strike. NATO doctrine already places it at the outset. AJP-3.9(B) states that “LEGADs should be involved in the targeting process from the beginning” (AJP-3.9(B), at 1.6.1, p. 1-23). It directs that IHL/LOAC principles be integrated at every stage, from a target’s identification through its validation to its engagement (same pinpoint). The baseline is therefore settled doctrine, not my contention: the adviser belongs in the process early and throughout.
The structure of the joint targeting cycle reflects this. The targeting staff (the multidisciplinary staff element that develops, validates and recommends targets to the commander) includes a legal adviser as a minimum member (AJP-3.9(B), at 1.5.2, pp. 1-21 to 1-22). Securing a legal assessment of targets and ensuring legal input into the process sit among its core functions (ibid.). Doctrine also locates advice at target development in Phase 2, and it requires formal records of the advice given, especially at Phases 2 and 4 (at Phase 2, pp. 1-15 to 1-16, and at 1.6, p. 1-23).
None of this is a recent discovery. The 1987 Commentary already described legal consultation as bearing on the preparation of plans, the choice of means and the determination of objectives (ICRC Commentary, at para. 3350). The early, planning-stage adviser is a feature of the role as it was understood almost forty years ago.
What Article 82 does not do is fix the point at which “early” begins. It leaves each State to determine when advice is necessary and at what level of command (Dinstein, Legal Advisers in the Field,at p. 919). That discretion is deliberate, and it is where AI-enabled operations make their mark. When planning and targeting draw on AI DSS, two things happen at once. The occasions on which advice is necessary multiply, as more decisions come to rest on system outputs. And the choices that settle lawfulness move forward in time. In operational terms, they lodge in the data a system is trained on, the parameters it runs under, and the validation it passes before an operation begins.
The practitioner reading is that this changes the content of “from the beginning,” not its existence. Being present early once meant being in the room when a target list took shape. It now also means being able to advise on choices made before that room convenes.
I do not claim that advisers are absent from these early choices. Doctrine says they should be there, and I take it at its word. The risk is narrower, and worth naming plainly. Involvement that satisfies the doctrine on paper may not yet reach the new substance of AI-enabled operations: the data, the parameters, the validation. That is a risk to manage through how the role is resourced and positioned. It is not a diagnosis that any State is failing today.
II. The duty was always elastic
Article 82 was written to flex. Its drafters built the give into the words themselves, and those words now do work they were never tested against in 1977. The provision requires that parties “shall ensure” advisers are available, then qualifies that duty twice: advice is owed “when necessary” and “at the appropriate level” (Article 82 AP I). Dinstein reads the clause as combining “a clear-cut obligation … with a certain degree of flexibility” (Dinstein, at p. 919). The obligation is firm; the conditions of its exercise are not.
That combination was deliberate. The 1987 Commentary records that “the obligatory character of the present provision was maintained,” even as the conditions of use were left open (ICRC Commentary, at para. 3344). The duty is mandatory in what it requires and flexible in how it is met. It also runs to the State. Article 82 obliges the party to the conflict to make advisers available (same paragraph), and it does not impose a personal performance standard on the individual adviser. That distinction carries weight later, when we separate what the adviser owes from what the State owes.
The drafters had their reasons for the flexibility. They did not want the validity of an order questioned merely because an adviser had gone unconsulted, nor the military hierarchy itself called into doubt (ICRC Commentary, at para. 3346). So the content of the obligation was allowed to vary from one State to another, according to the role each assigns its advisers. Some States appoint advisers at nearly all levels of command; others only at the headquarters of large units and at military academies; others again only in exceptional situations (at para. 3344). Article 82 accommodates these different arrangements by design, rather than prescribing one model.
Custom took the same course. The customary rule requires States to make advisers available when necessary, at the appropriate level (CIHL Rule 141). The elasticity is therefore no peculiarity of the Protocol. It is a feature of the rule at both the treaty and the customary level.
In my view, this built-in elasticity is what lets Article 82 meet AI-enabled operations without a new rule. The necessity trigger expands as those operations multiply the occasions that call for advice, as Section I set out. And the appropriate level travels to wherever the decisive choices are made, including the early, technical choices described earlier. No amendment is needed for the text to reach the new setting. It needs only to be read for what it already says.
That reading is textual, and I want to keep it modest. I am not suggesting that Article 82 speaks to AI, or that its drafters foresaw AI DSS. However, the conditions the drafters wrote to absorb variation across States absorb variation across decision environments too. A provision built to flex flexes here as well.
III. What “available” now requires
Being available is not the same as being present at the decision. Article 82 asks that advice reach the commander when necessary and at the appropriate level. In AI-enabled operations that asks two things of the adviser: to be positioned where the early choices are made, and to understand enough about the systems in play to advise on them. Neither is exotic. Both extend what the role already involves.
Positioning upstream
Doctrine already places the adviser along the whole targeting process, not at its end. AJP-3.9(B) casts the legal adviser as a subject-matter expert who reviews targets and carries IHL/LOAC through the cycle, from discovery to engagement (AJP-3.9(B), at 1.6.1, p. 1-23). Target validation is where compliance with the applicable legal framework is formally tested, and the adviser sits on the board that performs it (at Phase 2, p. 1-16, and at 1.3.33, p. 1-10). The adviser’s proper place already runs the length of the cycle.
The instinct to position the adviser early is not new either. Writing on the pre-deployment setting, Oakley argued that commanders should keep their advisers involved with current and future operations, an effort that should begin before forces mobilise (Oakley, Closing the Gaps, at “What to Train?”). That argument was about training. Its underlying logic, that useful legal advice attaches to preparation and not only to execution, is the logic Article 82’s “appropriate level” already carries, as Section II set out.
What AI-enabled operations change is how far upstream “across the cycle” now reaches. In operational terms, the choices that settle the lawfulness of later action increasingly sit in the data a system draws on, the parameters it runs under, and the validation it passes before deployment. Where those choices are, the appropriate level now extends. I should be plain that this is my reading rather than the doctrine’s: AJP-3.9(B) and Oakley locate the adviser early in the targeting and training cycles, but they do not address the design or validation of AI DSS. The extension to those stages is a practitioner inference, built on the positions the doctrine already assigns.
Availability also has a clock. Time-sensitive targets compress the standard targeting timeline, and AJP-3.9(B) lists timely access to a legal adviser among the conditions for engaging them successfully (AJP-3.9(B), at Annex A, A.2, p. A-1). The practitioner reading is that AI-enabled tempo sharpens this. As systems quicken the pace at which targets are found and engaged, the window in which advice can land narrows. Being positioned to be reachable inside that window is itself part of what availability now requires.
The literacy availability now demands
Positioning solves only half the problem. An adviser placed upstream must also understand enough about the systems to advise on them. NATO’s responsible-use principles require that AI applications be appropriately understandable and transparent, supported by verification and validation mechanisms (NATO AI Strategy (2021 summary), Principles of Responsible Use (here: explainability and traceability)). That understandability is what makes competent legal advice possible; without it, the adviser cannot test whether a proposed course of action is lawful.
Doctrine already expects the adviser to interrogate what is put in front of them. The legal adviser is an expert who makes specific requests for intelligence and assessments in order to advise the commander (AJP-3.9(B), at 1.6.1, p. 1-23). In operational terms, that habit of questioning carries directly into AI-enabled work. The adviser must hold enough of a working grasp of the systems in play to probe their limitations and to press the technical advisers on what those systems can and cannot support. That is sufficient understanding for the advisory task, well short of the inner-workings mastery of an engineer.
The practitioner reading is that explainability functions here as a precondition. It is what lets the adviser give correct advice on the lawful use of a system. It does not make the adviser responsible for that use. Who carries which responsibility is the question Section IV takes up.
IV. The responsibility the adviser carries
Responsibility for what AI DSS help produce stays with humans. It does not migrate to the system, however sophisticated the output. International humanitarian law binds States, parties to a conflict and individuals; it does not bind machines, and responsibility cannot pass to them (GGE Rolling Text (12 May 2025), at V(1)–(2)). NATO’s responsible-use principles say as much in the register of doctrine: the use of AI demands care and human judgement, and “clear human responsibility shall apply in order to ensure accountability” (NATO AI Strategy (2021 summary), Principles of Responsible Use (here: responsibility and accountability)).
The practitioner debate arrives at the same place. Writing on how the law of armed conflict absorbs AI, Geltzer accepts that a human must answer for what these systems do, and states plainly that “we cannot hold the AI model itself accountable” (Geltzer, A Case for Gradualism). He goes further, arguing that human accountability matters more as the technology enters weapons systems and military planning, and that the law is best served by holding onto it rather than letting it slip. That is the reverse of the accountability vacuum the critical literature fears.
None of this settles which human carries what. Here the adviser’s position has long been clear in kind, if not always in degree. The adviser advises; the commander decides. Dinstein describes the role as advisory only, with the decision left in the commander’s hands (Dinstein, Legal Advisers in the Field, at pp. 921–922). The 1987 Commentary marks the outer limit sharply: the adviser must never make an unlawful order appear lawful (ICRC Commentary, at para. 3367). The contribution is counsel; the authority to act on it or set it aside stays with the commander.
In my view, that division fixes what the adviser is answerable for in AI-enabled operations. The adviser is responsible for the correctness of the advice given, and for the adviser’s own part in ensuring that IHL is applied correctly. Responsibility for the decision rests with the commander who takes it. Responsibility for fielding the system rests with the State that deploys it. These are distinct answerabilities, and AI does not fuse them.
This responsibility also reaches back to availability. An adviser can answer for the correctness of advice only if positioned to give it and equipped to understand what the system offers. The positioning and literacy set out in Section III are therefore not separate from responsibility; they are its precondition. An adviser kept away from the early choices, or left unable to interrogate the system, cannot fairly be held to advice they were never in a position to give.
This is also where a familiar phrase misleads. The adviser is not the “human in the loop” in the sense the phrase usually carries, the human who watches a system’s processing and vouches for each step. That is not the adviser’s function, and the idea that lawful use requires a human to monitor every internal operation is one I have questioned before. The adviser is the human whose legal judgement the system is there to serve. The system informs the advice; the advice remains a human act of legal reasoning, attributable to the person who gives it. Whether that chain of accountability holds together across the phases of an AI-enabled operation is a question I have examined for the cyber-targeting setting, and I will not restate that analysis here.
The adviser’s responsibility, then, is real but bounded. It is the responsibility to advise well and to hold the line on the law. It is not the responsibility to command the operation, nor to field the system. That boundary is what keeps the adviser useful and keeps accountability where it belongs.
Conclusion
The legal adviser in AI-enabled operations does the work the law has always asked of the adviser. The task is to keep the commander’s decisions within IHL. What has changed is the setting in which that judgement is exercised, not the judgement itself.
Article 82 was written to accommodate that kind of change. Its firm duty sits alongside flexible triggers: advice when necessary, at the appropriate level. Those terms reach AI-enabled operations without amendment. The text was ready.
Because the text was ready, what the new setting asks for is institutional rather than legal. It asks that the adviser be positioned where the early choices are made, in the data, the parameters and the validation. It asks that the adviser understand enough of the systems to advise on them. And it asks that the adviser stay reachable when the timeline is short. None of this needs a new rule. It needs the function resourced to meet a more demanding environment.
Responsibility, meanwhile, stays where it has always sat. The adviser answers for the correctness of the advice and for the adviser’s part in seeing that IHL is applied. The commander answers for the decision. The State answers for the system it fields. AI does not blur these lines, and the adviser is not relieved of the duty to advise well because a system now informs the decision too.
The practitioner reading, then, is an optimistic one. The concern is a narrow and tractable one: that involvement adequate on paper may not yet reach the substance of AI-enabled operations. That is a task to manage, through how the role is positioned and resourced, and not a failure to allege.
My next post steps back from AI to revisit a foundational safeguard of AP I: the protection Article 41 extends to those who are hors de combat, the enemy who is surrendering, wounded, or otherwise out of the fight.