The Algorithm Is Not a Legal Adviser: AI Targeting and the IHL Duty of Human Judgment

AI can support military decision-making, but it cannot carry the legal judgment required by IHL. The hard question is not whether algorithms are useful; it is whether humans still understand, constrain, and remain responsible for the use of force.

Why it matters operationally

AI is attractive in warfare because it promises speed. It can process surveillance data, identify patterns, recommend targets, support counterfire, prioritize threats, and reduce some forms of human error. In high-tempo operations, especially against missile launchers, drones, air defenses, cyber infrastructure, or dispersed armed groups, commanders may see algorithmic tools as necessary rather than optional.

That is exactly why the legal question matters. IHL rules are applied before and during attacks, often under pressure and uncertainty. If an AI-enabled tool compresses the time between detection and strike, it may also compress the time available for legal assessment. If an autonomous weapon selects and engages targets after activation, the user may not know the precise person, object, time, or location that will be struck. If a decision-support system ranks thousands of possible targets, the human may be tempted to approve the output rather than interrogate it.

The career-relevant point for IHLBriefs is this: a serious legal/protection analyst should not simply say “AI is illegal” or “AI is just a tool.” Both statements are too crude. The better question is whether the system leaves humans with enough knowledge, time, control, and accountability to comply with IHL.

Facts and status

The ICRC describes autonomous weapon systems as weapons that, once activated, can select and engage targets without further human intervention. It emphasizes that the user may not know the specific target, timing, or location of the resulting force, and that IHL obligations must be fulfilled by humans rather than by the weapon system itself.

The CCW Group of Governmental Experts on lethal autonomous weapons systems has continued to discuss possible elements of an instrument or other measures on emerging technologies in the area of LAWS. The fact that the debate continues is itself important: there is broad agreement that IHL applies, but there is no full agreement on what additional rules or limits are required.

Applicable legal framework

This article concerns jus in bello, not jus ad bellum. It does not ask whether a state may use force under the UN Charter. It asks how IHL applies once an armed conflict exists and force is being used.

Key rules and concepts:

1. Distinction. Parties must distinguish between civilians and combatants and between civilian objects and military objectives. An AI system that cannot support reliable distinction in its intended environment creates a serious legal problem.

1. Military objective. For objects, the target must by its nature, location, purpose, or use make an effective contribution to military action, and its destruction, capture, or neutralization must offer a definite military advantage in the circumstances ruling at the time.

1. Proportionality. Even if the target is lawful, an attack is prohibited if expected incidental civilian harm would be excessive in relation to the concrete and direct military advantage anticipated.

1. Precautions. Parties must take feasible precautions to verify targets, choose means and methods that reduce civilian harm, and cancel or suspend attacks if the legal assessment changes.

1. Weapons review. For states bound by Additional Protocol I, Article 36 requires review of new weapons, means, or methods of warfare to determine whether their use would be prohibited in some or all circumstances. Even outside AP I treaty obligations, legal review is a serious compliance tool for AI-enabled capabilities.

1. Accountability. The presence of AI does not dissolve responsibility. A system that produces harm but leaves no human able to explain the assumptions, limits, training, data, and intended use creates legal and operational accountability problems.

The legal test: what human judgment must cover

Strongest argument for AI-enabled targeting

The strongest lawful-position argument is not that AI replaces legal judgment. It is that AI may improve the factual basis for legal judgment.

A well-designed system might detect incoming fire faster than humans, compare multiple data streams, reduce cognitive overload, flag civilian presence, improve target verification, or suggest lower-risk timing. In some environments, such as defensive counterfire against clearly military systems, a fast decision-support tool may reduce harm by enabling more precise and timely responses. A human commander with better information may make better IHL decisions.

This argument is strongest where the environment is constrained, the target category is clear, civilians are absent or sparse, the system has been legally reviewed, and humans understand the tool’s limitations.

Strongest legal concern

The strongest concern is not science fiction. It is ordinary legal judgment being degraded by speed, opacity, and automation bias.

If a system ranks people or objects as targets but the human reviewer cannot understand the basis for the ranking, proportionality and precautions become thin. If the human has seconds to approve a recommendation, the review may become formal rather than real. If commanders trust the tool because it appears technical and neutral, they may under-question bad data, biased assumptions, or stale intelligence. If an autonomous system searches for a generalized target profile in a changing environment, the human may not know enough about the specific strike to assess civilian harm.

The ICRC’s position is especially important here: it argues that IHL obligations must be fulfilled by humans and that new rules are urgently needed to clarify and specify how IHL applies to autonomous weapon systems. That does not mean every AI-related tool is unlawful. It does mean the law cannot be satisfied by saying “the algorithm decided.”

Analysis

The legal question should be broken into three different categories.

1. AI as information support

This is the least controversial category. AI may help process surveillance data, translate communications, detect patterns, or identify possible objects of interest. The legal risk depends on how the output is used.

If the tool only helps humans find information, and humans still verify the target and apply the legal rules, the system may be compatible with IHL. But even here, the review must be real. A commander cannot treat an AI recommendation as a legal conclusion. The tool may say, “this object resembles a launcher.” It cannot say, in any meaningful legal sense, “this attack is proportionate.”

2. AI as decision support

This is the hardest practical category. Decision-support systems may produce target lists, confidence scores, risk ratings, or strike recommendations. The system does not fire the weapon, but it shapes the human decision.

The legal risk is automation bias. Humans may over-trust the output because it appears objective. The risk increases when the target list is large, the tempo is high, the source data is classified or inaccessible, and the reviewer cannot test the assumptions. In such cases, a human signature at the end of the process does not necessarily equal human judgment.

For IHL compliance, the human must be able to ask and answer basic questions: What is the target? Why is it considered a military objective? What is the expected civilian harm? What assumptions does the system rely on? What is the error rate? What facts would require cancellation or delay? What alternative means or timing would reduce harm?

3. AI in autonomous weapon systems

This category raises the sharpest issue because the system may select and engage targets after activation. The human may set parameters but not choose each target.

Autonomy is not automatically unlawful. Some older weapons, such as certain mines or defensive systems, have long operated without real-time human control. But modern AI-enabled autonomy can operate in more complex environments and with less predictable behavior. The more open the environment, the more varied the possible targets, and the more civilians may be present, the harder it becomes for humans to make the legal judgments in advance.

The lawfulness of use therefore depends heavily on constraints: target type, geographical area, duration, environmental conditions, ability to deactivate, supervision, reliability, and the civilian density of the area. A system deployed against incoming missiles at sea is legally different from a system searching an urban neighborhood for people matching a pattern.

Protection implications

AI targeting is not only a weapons-law issue. It is a civilian-protection issue.

First, errors can scale. A bad human decision may cause one unlawful attack. A bad model or flawed dataset can replicate error across many decisions.

Second, speed can shrink protection space. Civilian presence, medical status, surrender, displacement, and protected objects are often context-dependent. If the decision cycle becomes too fast for those facts to be noticed, civilian protection weakens.

Third, accountability can become blurred. Was the harm caused by the developer, the commander, the operator, the intelligence analyst, the trainer, the data provider, or the state that deployed the tool? IHL responsibility ultimately remains with humans and parties to conflict, but practical accountability becomes harder when decision chains are opaque.

Fourth, field protection work may be affected. Humanitarian actors negotiating civilian movement, medical access, or protection of infrastructure need parties to understand and control their own systems. A party that cannot explain how a system selects targets will be harder to engage in meaningful protection dialogue.

What facts would change the analysis

The legal assessment would change depending on:

• whether the system merely supports information review or actually selects and engages targets;
• whether the system is used in a civilian-dense environment;
• whether target categories are narrow and objectively identifiable;
• whether humans can understand the system’s basis for output;
• whether commanders can suspend or deactivate the system;
• whether the system has been reviewed under Article 36 or an equivalent legal-review process;
• whether post-use audits reveal recurring civilian-harm patterns;
• whether operators receive training on the system’s limits and legal risks.

What cannot be concluded yet

This article cannot conclude that any specific state system is lawful or unlawful without technical and operational details. Public descriptions of AI tools are usually incomplete. Legal assessment requires information about design, training data, target profile, environment, human-machine interaction, command responsibility, testing, error rates, and actual use.

It also cannot resolve the broader diplomatic debate over whether new treaty rules should prohibit or regulate certain autonomous weapon systems. The better conclusion is that existing IHL provides the baseline, but the baseline may need clarification where autonomy weakens human control over life-and-death decisions.

Final assessment

The useful question is not whether the future of warfare will include AI. It already does. The useful question is whether legal judgment remains meaningful when machines accelerate the path from data to death.

For IHL purposes, the algorithm is not the legal adviser, the commander, the proportionality assessor, or the accountable decision-maker. AI can support human judgment only if the human still has enough information, control, time, and training to exercise judgment. When that chain breaks, the problem is not just technical. It is legal and humanitarian.

Sources used

Disclaimer: IHLBriefs is written in a personal capacity for educational and professional-portfolio purposes. It is not legal advice and does not represent the views of any university, employer, humanitarian organization, or institution with which the author is or has been affiliated.