The 90-Second War argued that AI literacy is now a personal and professional obligation for every ADF member. This article picks up where the previous one left off. When the loop is under 90 seconds and artificial intelligence (AI) has authored the recommendation, what standard is the commander held to when authoring a strike in an AI kill chain? In published Australian doctrine today, that standard has not yet been articulated. This article is about what we do until it is.
The 2026 National Defence Strategy is unambiguous. We are preparing for war.[i]
Strategic analysis suggests the United States (US) is setting the conditions in Venezuela and Iran by removing Chinese energy sources from the board before the war moves to the Pacific. The Center on Global Energy Policy at Columbia documents the disruption to Venezuela-China oil flows arising from US action;[ii] the Stimson Center catalogues the parallel pressures across Europe, the Middle East, and the Indo-Pacific;[iii] commentary at Modern Diplomacy reads the wider campaign as a deliberate, networked rollback rather than improvisation.[iv] The clock those operations imply is the one we must prepare for.
When the Australian Army deploys to that war, we will use the tactics, techniques, and procedures we have today. We will fight alongside our allies. We will answer to Australian law and to the Australian people, from whom the legitimacy of every strike, every detention, every act under arms is drawn. The two obligations are compatible. They are not interchangeable. A coalition partner’s targeting standard is a coalition partner’s. The Australian commander in the seat answers to a different authority.
Today, that commander must exercise command responsibility for AI-derived target recommendations they neither built nor reviewed, and cannot fully audit, on a clock that closes in ninety seconds. The Defence AI Policy of March 2026 sets out the principles.[v] It does not yet define the operational standard. Defence-level doctrine continues to be reviewed. None of that work addresses the deployment timeline.
We have a brief window to prepare our junior leaders and operational commanders for the moment, as the doctrine has not yet been built to meet it. The Cove exists to foster professional discussion. What follows is one operational moment, then a reading of the public record, and a proposal for what the institution and each reader can do before the next operational deployment.
The recommendation arrives at 0247.
The model is partner-provided. The target is one of the seven on the dynamic list that the Combined Joint Operations Centre has been working on since the previous shift. The pattern of life has been built by a partner J2 cell over sixteen hours, none of whom you have met. Your Legal Officer is on the radio beside you. The partner Legal Officer is on the second radio. The window to authorise is ninety seconds. Confidence point nine four. Your Legal Officer only has thirty seconds to provide their advice.
The principle of distinction applies. The principle of proportionality applies. The principle of precaution applies. None of them, in published Australian doctrine, has a settled application to the recommendation on your screen.
Your training did not provide you with a standard operating procedure (SOP) for this moment.
You invent one. You key the radio beside you. You ask the question your lead-up training did not prepare you for. Are we operating to their standard or ours?
The Legal Officer is silent for two seconds. He answers. You key the second radio. You verify with the partner J2 by posing a question you compose in real time. The clock continues.
You authorise. The strike is clean. The collateral damage estimate holds. You return to your room.
You think about it for six months after.
The question that does not leave is this. Under what frame did you answer in the ninety-second window, when the standard had not been articulated, the doctrine had not been written, and the operating procedure had not been explained?
The answer the institution has given you is the only one available under Australian law.
You were lucky that the process you invented in the chair was the right one. The next operational commander in that chair might not be.
The Defence AI Policy of March 2026 is clear about Article 36. It commits Australian Defence to three policy requirements: lawfulness (under domestic and international law, including Article 36), adherence to values-based principles, and proportionate controls. It is the right policy. However, according to The Conversation's March 2026 analysis, it is thin on implementation.[vi]
The doctrinal flow-down that translates these principles into an operational standard continues to be reviewed. ADF-I-3 Targeting and ADF-I-0 Law of Armed Conflict, the foundational ADF doctrines on lawful targeting, are the natural homes for the AI-specific procedural architecture a commander would need at the targeting board. Neither closes the operational gap this article addresses: what standard the soldier is held to when the recommendation is AI-derived, the tool is the coalition partner’s, and the window to authorise is measured in seconds.
What this article adds is the public-record diagnostic. It draws on what Australia’s coalition partners and others have already published on AI targeting, including the US, UK, NATO, and Israel. It synthesises legal, ethical, doctrinal, and operational scholarship from outside the chain of command. And it closes with five unit-level exemplars: draft procedures that invite the SOP work the formal doctrine cycle has yet to release.
Three policy formulations, one legal substrate
Three policy formulations are in operation across Australia’s coalition partners.
The US operates under the principle of ‘appropriate levels of human judgment over the use of force’, as set out in the Department of Defense Directive 3000.09, reissued on 25 January 2023. The directive requires that autonomous and semi-autonomous weapon systems be designed to allow commanders and operators to exercise that judgement.[vii]
The United Kingdom (UK) operates under ‘appropriate human oversight’, per Joint Service Publication 936 V1.1 (13 November 2024).[viii] The wider UK Ministry of Defence doctrine framework around JSP 936, set out in Ambitious, safe, responsible, articulates the position as ‘meaningful human control… through context-appropriate human involvement’.[ix]
NATO operates to ‘appropriate levels of judgment and care; clear human responsibility shall apply’, per the Principles of Responsible Use of AI in Defence (October 2021, revised July 2024).[x]
Forty-two states, in a September 2025 Joint Statement to the Group of Governmental Experts (GGE) on Lethal Autonomous Weapons Systems, advocate ‘meaningful human control’.[xi] Australia is not among them. Australia’s GGE position is permissive in form and AUKUS-aligned in substance, as reflected in working papers it has co-sponsored with the US and the UK.
Meaningful human control is a substantive legal proposal with concrete implications for procurement, training, and authorisation. It is not a slogan. This article declines to advocate that Australia adopt it as a stand-alone doctrinal standard. The procedural architecture proposed in the closing is consistent with its intent but operationalises it differently for AUKUS coalition-interoperability.
The legal standard common to all three formulations is Additional Protocol I and customary international humanitarian law. The reasonable commander test still applies. Distinction. Proportionality. Precaution. What the formulations differ on is the procedural architecture, the training expectations, and the audit posture each implies for the human in the chair. They are siblings, not three crisp alternatives.
Australia has not chosen between them in published doctrine. In practice, Australian commanders deployed alongside coalition partners operate to whichever standard the partner has built into the targeting tool. The Australian articulation that tells the commander which architecture is consistent with our diplomatic and doctrinal posture is in the works.
This article will, in its closing, return to a procedural architecture sharper than the current diplomatic posture would comfortably accommodate. That tension is named here. The architecture is calibrated to the institutional good-practice question, not the diplomatic-alignment question.
The Israeli Defense Force’s (IDF's) use of Lavender, Gospel, and Habsora in Gaza is the live case study of what happens when the procedural architecture is left to operator discretion at scale.
Reporting by Yuval Abraham in +972 Magazine and Local Call in April 2024 described large-scale institutional reliance on AI-derived target lists, operator authorisation cycles compressed below traditional verification thresholds, and contested questions about the rules governing authorisation.[xii] The report drew on six unnamed Israeli intelligence officers. The specifics, including operator-window figures, error-rate estimates, and target-volume claims, are disputed.
The IDF response, issued through the Spokesperson Unit in April-May 2024, acknowledged the existence of Lavender. It characterised the system as ‘a database whose purpose is to cross-reference intelligence sources’ and rejected the autonomous-system reading. It stated that the systems are ‘designed to aid human analysis, not to replace it’ and described the +972 figures as ‘baseless in fact’ and reflecting ‘a flawed understanding of IDF directives and international law’.[xiii] The exchange has been engaged substantively in the academic literature, including the Lieber Institute (West Point) symposium on Gospel, Lavender and the Law of Armed Conflict, and Just Security commentary across 2024 and 2025.[xiv][xv]
The structural point survives the dispute. Whether the operator window was 20 seconds or 200, whether the error rate was 10 per cent or 1 per cent, or whether the target list was 37,000 or 3,700, an institution that fields AI-derived target recommendations at scale requires an articulated procedural architecture for how the human authority is discharged. The Lavender, Gospel, and Habsora episode is the live case study because the architecture was either inadequate to the tempo, on the +972 reading, or adequate but undocumented in a way that cannot withstand independent review, on the IDF reading. Either way, the structural question is one the institution can no longer leave unanswered.
The procedural architecture is one half of the problem. The cognitive architecture is the other. Peter Layton, writing for RUSI in January 2025, identifies what the Lavender episode reveals about how human-machine teams actually think.[xvi] The team’s cognition is shared, and the machine’s logic reshapes human cognition. AI privileges what is quantifiable over what is qualitative; it favours the easier target over the harder one; it encourages, in Layton’s phrase, automation bias and action bias in the human partner. IDF personnel reportedly approved Lavender recommendations after twenty seconds’ consideration and treated the system’s outputs as equivalent to human analysis. That is a cognitive failure.
The institution that deploys AI-driven targeting at scale must maintain two architectures, not one. The procedural architecture sets out how the human authority is discharged. The cognitive architecture sets out how the human’s judgement is protected from the bias that the system itself encourages. The reform vectors and exemplars proposed later in this article address both, but the cognitive question deserves to be named on its own terms. The institution that asks only ‘did the operator follow the procedure’ without also asking ‘did the operator still think clearly under the recommendation’s gravitational pull’ will get the procedural process right and the cognitive question wrong.
The Australian intellectual heritage
The Australian institutional vocabulary for the question already exists, but it has not yet been translated into a published doctrinal standard against which a commander can be trained.
Australian defence research at Trusted Autonomous Systems Defence CRC and UNSW-Canberra developed the ‘minimally-just autonomy’ frame in AI in Weapons: The Moral Imperative for Minimally-Just Autonomy (ICSILP 2018).[xvii] The frame articulates, in Australian terms, an ethical floor for autonomous targeting: the protected objects and surrender signals a system must learn to avoid. The Australian Army Journal has published two anchor pieces: ‘Autonomous weapon systems, international law and meaningful human control’ (Vol 13 No 1) and ‘Operational AI Integration and Governance in the Australian Army’ (Vol 21 No 1).[xviii][xix] The Australian Army Research Centre’s Land Power Forum piece, ‘Mission Command and Artificial Intelligence: Obstacles to Integration’, identifies the integration problem from the inside.[xx]
A coalition partner officer at a recent leadership course described the Australian frame this way. A constitutional answer, mediated by a specific law, in which the commander in the seat is the institutional point at which the constitutional bargain meets the trigger. It is a heavier answer than our coalition partners’. It is also, distinctively, ours.
The intellectual material exists. The doctrinal product is in production. This article is the bridge.
The diagnostic ends here. The proposal that follows is offered as an institutional contribution to a doctrinal cycle already underway. The reader is invited to test the diagnostic and the proposal against their own work area. What the institution can do and what each reader can do are different scales of the same question.
Three institutional reform vectors
The article proposes a procedural and training architecture that supports the work and the commander in the meantime. Three vectors are worth Defence’s consideration.
A contextual articulation framework, published at the appropriate classification. Defence should publish the matrix that operationalises the standard: targeting phase (intelligence, surveillance and reconnaissance, target development, target acquisition, target engagement, effects assessment) by tempo (deliberate, dynamic, time-sensitive) by AI role (decision support, recommendation, recommendation with execution). Each cell of the matrix becomes a doctrine line for an articulated procedural standard. The article does not propose what fits in each cell. It proposes that the matrix exist, be published at the appropriate level, and serve as the framework against which Capability Manager instructions and Service-level doctrine are written.
An Article 36 in-service re-review trigger for AI-DSS. The Defence AI Policy commits to an Article 36 review at acquisition. The live academic question, advanced in the work of Klaudia Klonowska on AI-DSS Article 36 review and Damian Copeland on AI as a method of warfare, is whether AI decision-support systems require in-service re-review when the underlying model is updated, retrained, or replaced.[xxi][xxii]
The Anthropic-US Department of Defense episode from February to April 2026 is the live case study of what happens when the AI vendor changes mid-doctrine. On 27 February 2026, a US executive order terminated the Anthropic federal contract; on 26 March 2026, the Northern District of California issued a preliminary injunction (Judge Rita Lin); on 8 April 2026, the DC Circuit denied Anthropic’s emergency stay request.[xxiii] Defence publishing an in-service re-review trigger for AI-DSS, as part of the next tranche of doctrinal flow-down, would be a useful step for the in-train cycle to take.
An Australian Defence College (ADC) / Land Combat College (LCC) curriculum module on AI-derived targeting authority. The professional military education pipeline is the slowest variable in the system. A curriculum module at the officer entry and staff college levels, taught against the contextual articulation framework once published and, in the meantime, against the existing Defence AI Policy and the customary command responsibility doctrine of Additional Protocol I Articles 86 and 87, would put the standard into the cohort of officers most likely to be in the chair when the moment arrives.[xxiv] An officer at the same leadership course put the urgency simply: the doctrine cycle is slower than the enemy's. The training cycle is even slower. Starting the curriculum work now is the only way the cycle can catch up before the deployment.
Five exemplars to provoke unit-level SOP work
We are sending junior leaders and operational commanders into the next war with a procedure each will invent in the chair, and we will find out later whether they invented the right one. The exemplars below show one shape the work of inventing those procedures could take. The SOP work belongs in the unit, the formation, and the targeting cell. The article presents exemplars to provoke that work.
The AI confidence call-and-response. A verbal protocol at the targeting board: the operator states the recommendation, confidence score, model identifier, and data provenance flag in a single utterance; the commander acknowledges each element by name before authorisation. The unit’s SOP would specify the wording, the position in the F2T2EA targeting cycle (Find, Fix, Track, Target, Engage, Assess), the named authority, and the documentation artefact. A time-sensitive variant would shorten the verbal protocol while preserving the four elements as a checklist for the commander to sign.
The bias-flag protocol. A tagging convention for AI-recommended targets when the underlying training data is suspect: provenance unknown, partner-provided, or the model has been recently updated. The flag does not stop the strike. It changes the documentation requirement and the after-action review trigger. The unit’s SOP would specify the named flagging authority, the flag categories, the artefact form, and the after-action review threshold.
The second pair of eyes rule. A second-authority requirement for AI-recommended strikes above a defined threshold: collateral damage estimate category, civilian harm probability, or a partner-provided recommendation. The unit’s SOP would specify the threshold, the rank and role of the second authority, the maximum delay introduced, and the authorisation artefact.
The Legal Officer authorisation script. A five-question protocol that the Legal Officer runs in real time, and the commander acknowledges. Example questions, calibrated to the AI-derived targeting context and to be refined by the unit for its mission set:
- Is the target distinguishable as a lawful target on the verification we hold, independent of the AI?
- Does the proportionality assessment hold against the collateral damage estimate?
- What is the data provenance of the recommendation, and which flags apply (partner-provided, model recently updated, training-data unverified)?
- Are we authorised under Australian rules of engagement on this strike, regardless of the partner standard the tool reflects?
- If you had ten more seconds, what would you ask before authorisation?
The script supplements, rather than replaces, the Legal Officer’s professional judgement. The unit’s standing operating procedure would specify the script’s authority, the documentation form, and the local refinements.
The AI-derived strike after-action review trigger. An after-action review record at the unit level that names the model, the confidence, the verification artefacts, and the training-data flag. The records aggregate to a formation-level review and feed back into doctrine. The legal anchor for the distributed-accountability frame is the customary command responsibility doctrine under Additional Protocol I Articles 86 and 87. The institutional good-practice anchor is the unit’s existing after-action review culture. The unit’s SOP would specify the named review convening authority, the artefact retention rule, the formation-level aggregation cadence, and the doctrine-cycle feedback path.
These five are exemplars only. The article’s prompt to the reader is to choose which of these your unit would trial before the next exercise, and to describe what shape the standing operating procedure would take in your context.
The Defence AI Policy sets out the principles for Australian commanders. The doctrinal articulation that translates those principles into the standard a commander is held to in dynamic targeting at coalition tempo is the institution’s work, and that work is underway. While it is underway, three prompts are worth keeping in the reader’s work area this week.
First, identify one AI-enabled targeting tool your area has fielded or will field by 2028. Name the policy formulation your coalition partner is operating under. In writing, identify the procedure your unit follows when the recommendation arrives.
Second, identify one of the five exemplars that this article proposes for your targeting cell, firing cell, or planning cycle to trial before the next exercise. The cost is low, and the artefact it generates feeds back into the doctrine cycle.
Third, take both to the next professional discussion in your chain of command. The Cove can name the gap that the doctrine cycle is closing. In the meantime, closing it is the work of every targeting board, every Legal Officer brief, every after-action review, and every reader.
Under what frame will you answer, when the commander in the seat is you, the window is sub-90 seconds, the standard is undefined, and the people you answer to are the Australian people from whom the legitimacy of the answer is drawn?
As someone who has been in the room and still has unanswered questions, I remind you that Courage and Integrity are core ADF values for a reason. You can say no.
Still Interested?
Why not read this article by Christopher Taylor, titled ‘Leveraging Artificial Intelligence in ISTAR Operations: Opportunities and Challenges’.
Did you also know Cove+ offers a course on Information Technology: Artificial Intelligence?
End Notes
[i] Australian Government, Department of Defence (2026). 2026 National Defence Strategy. April. https://www.defence.gov.au/about/strategic-planning/2026-national-defen…
[ii] Center on Global Energy Policy, Columbia University SIPA (2026). US Action Threatens Venezuela-China Oil Flows, Debt Repayment, and Investments. https://www.energypolicy.columbia.edu/venezuela-china-oil-ties-severely…
[iii] Stimson Center (2026). Top Ten Global Risks for 2026. https://www.stimson.org/2026/top-ten-global-risks-for-2026/
[iv] Modern Diplomacy (2026). Not Chaos but Grand Strategic Design: The Hidden Logic Behind Washington’s Recent Moves. 25 March. https://moderndiplomacy.eu/2026/03/25/not-chaos-but-grand-strategic-des…
[v] Australian Government, Department of Defence (2026). Policy Settings for the Responsible Use of Artificial Intelligence in Defence. March. https://www.defence.gov.au/sites/default/files/2026-03/Policy-Settings-…. The policy states at page 7: “Australia is required by Article 36 of Additional Protocol 1 to the Geneva Conventions to conduct legal reviews of all new weapons, means and methods of warfare to ensure that their employment in some, or all circumstances, is lawful.” The three policy requirements are stated at page 6 as: (1) Lawfulness; (2) Adherence to values-based principles; (3) Proportionate controls.
[vi] The Conversation (2026). Australia’s new military AI policy comes at a crucial time. The challenge is turning it into practice. March. https://theconversation.com/australias-new-military-ai-policy-comes-at-…
[vii] United States Department of Defense (2023). DoD Directive 3000.09 – Autonomy in Weapon Systems. Reissued 25 January. https://www.esd.whs.mil/portals/54/documents/dd/issuances/dodd/300009p…
[viii] United Kingdom Ministry of Defence (2024). Joint Service Publication 936 V1.1: Dependable Artificial Intelligence in Defence (Part 1: Directive). 13 November. https://assets.publishing.service.gov.uk/media/6735fc89f6920bfb5abc7b62…
[ix] United Kingdom Ministry of Defence (2022). Ambitious, safe, responsible: our approach to the delivery of AI-enabled capability in Defence. https://www.gov.uk/government/publications/ambitious-safe-responsible-o…
[x] NATO (2024). Summary of NATO’s Revised Artificial Intelligence Strategy. 10 July. https://www.nato.int/en/about-us/official-texts-and-resources/official-…
[xi] Group of Governmental Experts on Emerging Technologies in the Area of Lethal Autonomous Weapons Systems (2025). Joint Statement of 39 High Contracting Parties and 3 Observer States. September. https://www.stopkillerrobots.org/news/september-2025-gge-joint-statemen…
[xii] Abraham, Y. (2024). Lavender: The AI machine directing Israel’s bombing spree in Gaza. +972 Magazine and Local Call, April. https://www.972mag.com/lavender-ai-israeli-army-gaza/
[xiii] Israeli Defense Forces Spokesperson Unit (2024). Public statements on Lavender, Gospel and Habsora, April-May. Quotes attributed via secondary reporting at Fortune (4 April 2024): https://fortune.com/2024/04/04/israel-idf-ai-warfare-gaza-palestinian-h…
[xiv] Lieber Institute, West Point (2024). Gospel, Lavender and the Law of Armed Conflict symposium. https://lieber.westpoint.edu/gospel-lavender-law-armed-conflict/
[xv] Just Security commentary series, 2024-2025. Unhuman Killings: AI and Civilian Harm in Gaza (March 2024): https://www.justsecurity.org/90676/unhuman-killings-ai-and-civilian-har…. Rethinking Responsible Use of Military AI: From Principles to Practice by Bridgeman, T. and Rosen, B. (26 September 2024): https://www.justsecurity.org/100378/military-ai/
[xvi] Layton, P. (2025). Human-Machine Teaming’s Shared Cognition Changes How War is Made. RUSI Commentary, 27 January. https://www.rusi.org/explore-our-research/publications/commentary/human…
[xvii] Scholz, J. and Galliott, J. (2018). Artificial Intelligence in Weapons: The Moral Imperative for Minimally-Just Autonomy. International Conference on Science and Innovation in Land Power 2018, Defence Science and Technology Group. https://www.dst.defence.gov.au/sites/default/files/basic_pages/documents/ICSILP18Wed1600_Scholz_et_al-AI_in_Weapons.pdf
[xviii] Kua, C. (2016). Autonomous weapon systems, international law and meaningful human control. Australian Army Journal Vol 13 No 1 (Autumn). https://researchcentre.army.gov.au/library/australian-army-journal-aaj/…
[xix] Wood, B. J. (2025). Operational AI Integration and Governance in the Australian Army. Australian Army Journal Vol 21 No 1. https://researchcentre.army.gov.au/library/australian-army-journal-aaj/…
[xx] Huang, K. Y. (2021). Mission Command and Artificial Intelligence: Obstacles to Integration. Land Power Forum, Australian Army Research Centre, 1 July. https://researchcentre.army.gov.au/library/land-power-forum/mission-com…
[xxi] Klonowska, K. (2020). Article 36: Review of AI Decision-Support Systems and Other Emerging Technologies of Warfare. Yearbook of International Humanitarian Law, Vol 23. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3823881
[xxii] Copeland, D. The Article 36 legal review of weapons enhanced by AI (autonomous weapons). Australian National University Open Research Repository. https://openresearch-repository.anu.edu.au/items/7d83b7ca-fa81-4b75-944…
[xxiii] Coverage of the Anthropic-United States Department of Defense dispute. The 26 March 2026 preliminary injunction (Northern District of California, Judge Rita Lin) is reported at https://www.cnbc.com/2026/03/26/anthropic-pentagon-dod-claude-court-rul…. The 8 April 2026 DC Circuit denial of Anthropic’s emergency stay is reported at https://www.cnbc.com/2026/04/08/anthropic-pentagon-court-ruling-supply-… and analysed at https://www.joneswalker.com/en/insights/blogs/ai-law-blog/two-courts-tw…. Wikipedia overview: https://en.wikipedia.org/wiki/Anthropic%E2%80%93United_States_Departmen…
[xxiv] International Committee of the Red Cross. Customary International Humanitarian Law, Rules 152 and 153 (the principle of command responsibility). Articles 86 and 87 of Additional Protocol I to the Geneva Conventions (1977). https://ihl-databases.icrc.org/en/customary-ihl/v1
