Accountability at the Speed of the Machine

Saturday, 24 October 2026

An AI system does not have to fire a weapon to change who is at risk. It can sort a thousand sensor reports into a list, put one person or building at the top, and attach a score that looks more certain than the evidence beneath it. A commander may still approve an attack. But the system has already helped decide what the commander sees first, what gets checked, and how long there is to challenge it.

This is the central problem in military AI: decision power can move through a chain without any single step appearing to surrender control. To see where responsibility lies, separate three things that are often collapsed into one headline: an algorithmic recommendation, the selection of a target, and the actuation of force.

From a recommendation to an attack

An AI-enabled decision-support system can gather and summarize satellite images, sensor readings, communications data or prior reports. It may identify patterns, rank options, estimate risk or recommend where to investigate. It is a tool in a human decision process; its recommendation is not itself an order to attack. The International Committee of the Red Cross (ICRC) says such systems can help analyze information, but warns that they can also affect decisions that lead to the use of force. ICRC, decision-support FAQ, 8 October 2026

The next step is target selection: a person decides whether the recommended object or person is a lawful military objective in the actual circumstances. That calls for context, not simply a model’s classification. Then comes the decision and means to apply force. A weapon can be autonomous in flight or navigation without being autonomous in target selection. Conversely, a weapon that, once activated, selects and engages targets without further human intervention raises a different set of legal and humanitarian questions. The ICRC makes this distinction explicitly: not all armed drones are autonomous weapons, not all autonomous weapons use AI, and AI in a weapon’s navigation does not itself make the weapon autonomous. ICRC, autonomous-weapons FAQ

The 2026 UN Group of Governmental Experts on lethal autonomous weapon systems used a similar functional line in its September report: the system’s integrated components identify, select and engage a target causing loss of life without intervention by a human operating it. The report also states that international humanitarian law (IHL) applies, that obligations bind people and states rather than machines, and that humans remain responsible for decisions involving these systems. This was a consensus set of elements for a possible instrument, “without prejudging its nature”; it was not itself a new treaty. UN CCW GGE report, 14 September 2026

A score can compress time

Military units already face more data than people can inspect one item at a time. An AI system can sift information faster, which may help a staff find a real threat in a crowded picture. Yet the same speed can increase the number of alerts, shorten the time available to check them and encourage staff to accept a machine’s ranking as a kind of answer.

The UNIDIR and ICRC literature calls one danger automation bias: people can over-trust a computer output, especially under pressure, fatigue or time constraints. There is an opposite failure too: distrust of a tool that is useful. Neither blind acceptance nor reflexive rejection is sound judgment. Good human control requires the operator to understand what the tool measured, what it omitted, how uncertain its output is and whether independent evidence agrees. UNIDIR, military AI report, July 2025 · ICRC, decision-support FAQ, 8 October 2026

Consider a deliberately simple example. Imagine a screening system reviews 10,000 reports, and just 1 per cent contain the indicator it is meant to find. A toy classifier catches 90 per cent of those 100 cases and correctly clears 99 per cent of the other 9,900. It produces 90 true alerts and 99 false ones: more than half of its 189 alerts are false. This is base-rate arithmetic for illustration, not an estimate of any weapon or named system’s performance. In the field, the true prevalence, model accuracy, data quality and consequence of an alert all differ. The lesson is narrower: a low false-positive rate can still produce a large number of false alarms when the underlying event is rare.

If the model’s output is shown as a confident ranked list, a human may then focus on the false alert and seek evidence that confirms it. Time pressure magnifies that risk: a commander might have less time to ask whether a communication was misread, whether the person has changed location, or whether an ordinary civilian pattern was treated as hostile. A peer-reviewed study of targeting support discusses how ranked outputs and time pressure can weaken validation and proportionality judgments. International Review of the Red Cross, December 2025 An adversary may also manipulate the data environment. The ICRC specifically warns that poor or context-insensitive data and deliberate manipulation can undermine decision-support outputs in armed conflict. ICRC, decision-support FAQ, 8 October 2026

Mistaken attribution is one possible route to escalation. A model might flag an event as consistent with an adversary’s action; that prediction is not proof of who caused it or why. If commanders act on the score before independent review, the system can accelerate a mistaken inference into a military response. The problem need not be an autonomous machine deciding to wage war. A human can authorize force while the machine has narrowed the alternatives, supplied a faulty premise and compressed the time for correction.

Logistics also carries human consequences

Military AI is not confined to weapons. It may help estimate equipment failures, manage fuel and spare parts, model transport routes or allocate staff. UNIDIR has mapped applications ranging from logistics and maintenance to command-and-control support; its October 2026 note also describes industry choices around data, simulation and autonomy thresholds. In these settings, good forecasts can reduce wasted supplies, identify bottlenecks and help a force keep essential services running. UNIDIR, Artificial Intelligence Beyond Weapons, 11 October 2023 · SIPRI, military AI industry map, 23 April 2026 · UNIDIR, 22 October 2026

But logistics decisions can move harm even when they do not choose a target. A supply model optimized only for mission readiness may direct scarce fuel or repair capacity away from routes used by civilians or humanitarian teams. A maintenance forecast based on incomplete records may miss equipment needed for evacuation or demining. These are conditional risks, not claims that a named military has deployed such a system or caused a specific outcome. They show why the system’s objective matters: “reduce delay” or “maximize availability” still means someone chose what gets prioritized.

Meaningful control is more than a human click

Human control is not just a person appearing at the final approval screen. A person cannot exercise meaningful judgment if the system’s limits are hidden, the data are stale, the interface buries uncertainty, or the time allowance makes independent checks impossible. Nor is an operator’s ability to press “stop” enough if they lack authority to do so or face incentives to accept every recommendation.

The OECD’s general framework for classifying AI describes a range of action autonomy: from a tool that merely supports a human’s choices, through systems that act with approval or under human supervision, to systems that act without human involvement. This is not a military legal category, but it helps make the interface visible: who can reject an output, who decides the target, and who or what triggers the action? The OECD’s broader AI Principles call for human agency, oversight and override safeguards as well. OECD Framework for the Classification of AI Systems · OECD AI Principles, updated 2024

For weapons that use autonomous functions, the UN’s September 2026 consensus text says control measures should take account of operational context and the system’s capabilities. It lists restrictions on scale, target types, duration and geographic scope, as appropriate; intervention and timely deactivation; realistic testing; clear training; mitigation of algorithmic and automation bias; and an accountable chain of command. These are elements to consider at the UN’s November review conference, not a settled international rulebook. UN CCW GGE report, September 2026

National policies have also begun to define lifecycle controls. The UK’s defence AI directive calls for governance, testing, safety and security across development and use. The US Department of Defense requires review of certain autonomous weapon systems and says the system, its interface, doctrine and training should let commanders and operators exercise appropriate human judgment over force. Such policies show that defence institutions can set controls; they do not establish how consistently every deployed system meets them. UK Ministry of Defence, JSP 936, 13 November 2024 · US DoD Directive 3000.09, 25 January 2023

What civilian protection requires

IHL requires distinction between military objectives and civilians or civilian objects, proportionality in expected incidental harm, and feasible precautions. Those are judgments for responsible human decision-makers. An algorithm can help gather data on civilian presence, compare options or flag uncertainty; it cannot turn a correlation score into a legal finding. The ICRC’s October 2026 recommendations call for reliable data, testing and legal reviews, trained users, meaningful engagement and the ability to challenge outputs. It also says certain uses should be prohibited, including autonomous weapons that target people directly or have unpredictable effects. These are the ICRC’s policy positions, not a claim that every state has adopted them. ICRC, decision-support FAQ, 8 October 2026

There is a serious countercase. Well-tested systems may help analysts inspect more information, find patterns they would otherwise miss and identify civilian locations or infrastructure for review. They can support decisions to avoid civilian harm if the data are relevant, the assumptions are clear and commanders have time to question the result. ICRC notes that responsible AI decision support can, in some circumstances, provide more comprehensive analysis that supports IHL compliance. The safeguard is not to reject useful tools; it is to test them in realistic conditions, disclose uncertainty, review failures and preserve independent human assessment.

The practical line is simple, even when the technology is not: a recommendation can inform a decision; it cannot inherit the responsibility for it. A human who is rushed, misled or unable to intervene may be present but not meaningfully in control. In war, that difference is not a philosophical footnote. It is part of how civilians are distinguished, protected and held in view when the screen fills with targets, routes and probabilities.

Evidence cutoff: 24 October 2026. The UN consensus elements cited above are non-binding pending consideration by the CCW Review Conference.

U.S. Navy watch standers review a battle assessment during a 2005 exercise aboard USS Mount Whitney. Historical file photo illustrating human review in operational decisions. Credit: U.S. Navy photo by Photographer’s Mate 2nd Class Greg Roberts, via Wikimedia Commons. Image source · Public domain in the United States as a U.S. Navy work made in official duties