AI on the battlefield: targeting, guidance and ethics

Military artificial intelligence did not arrive on the battlefield as an armed robot. It arrived as an image classifier. Its first documented job was to look: to work through hours of surveillance video that no human team could realistically review. From there it moved, step by step, into sensor fusion, precision guidance support and the drafting of proposed target lists. That trajectory is what turned AI targeting from an engineering question into a legal one.

This article sets out what military AI demonstrably does in publicly acknowledged programmes, which rules apply to it, and what it changes for accountability, decision speed and deterrence. It deliberately avoids the internal workings of any specific system: the aim is to understand the argument, not to describe how anything is operated.

Guided weapons came first, machine learning came later

Precision guidance predates modern AI and does not depend on it. The laser-guided bombs of Vietnam and the satellite-guided munitions of the 1990s work through sensors and classical control theory, with no machine learning involved. What today's AI adds is not marksmanship but information handling: recognising patterns in imagery, spotting what changed on the ground between two satellite passes, sorting vehicle types, ranking alerts among thousands of detections.

The most frequently cited public case is Project Maven, established by the US Department of Defense in 2017 under the formal name Algorithmic Warfare Cross-Functional Team. Its initial task was to apply computer vision to drone video in order to detect and label objects. In 2018 it triggered an internal revolt at Google and the company's withdrawal from the contract — the episode that opened the public argument about the tech industry's role in defence work.

Target selection means decision support, not automatic decisions

In every publicly described programme, "AI targeting" refers to decision-support tools: systems that pull together sensor feeds, intercepted communications and archived files, then hand a commander a list of proposed targets with the reasoning attached. Under the declared procedures of the forces that use them, approval remains human.

That arrangement has been the subject of contested reporting. Israeli and international outlets reported in 2023 and 2024 on target-recommendation tools used by the Israel Defense Forces in Gaza, naming systems such as Habsora (The Gospel). The IDF responded that these are analytical aids subject to human review and denied that they authorise strikes on their own. The available material is journalistic, and neither the figures nor the internal workings have been independently verified, so it belongs here as an open controversy rather than an established fact.

The deeper concern raised by lawyers and humanitarian organisations is not that a machine pulls the trigger. It is automation bias: when a system proposes hundreds of targets an hour, human review can decay into formal sign-off, with no real time to examine each case. The signature stays human; the deliberation may not.

The rules that already apply

There is no dedicated treaty on autonomous weapons, but there is no legal vacuum either. International humanitarian law applies, alongside national directives and multilateral processes:

• International humanitarian law: distinction, proportionality and precaution bind the attacker regardless of the technology used.
• US Department of Defense Directive 3000.09, issued in 2012 and reissued in January 2023, requiring appropriate levels of human judgement over the use of force and formal review of systems with autonomous functions.
• The Group of Governmental Experts under the Convention on Certain Conventional Weapons, debating lethal autonomous weapon systems since 2014 without producing a binding instrument.
• The International Committee of the Red Cross, which called for new rules and specific prohibitions in 2021 and joined a 2023 appeal with the UN Secretary-General for a legally binding instrument.
• The UN General Assembly, which adopted its first dedicated resolution on lethal autonomous weapon systems in December 2023, procedural in nature.

What "human control" actually means

"Meaningful human control" captures the central demand of the debate, and its vagueness explains why states cannot agree. Between full supervision and full autonomy lie intermediate settings that different militaries describe in different language.

Automatic defences against missiles and rockets, in service for decades, already sit in the first two categories with minimal intervention — precisely because the threat arrives faster than an operator can react. The disagreement is concentrated on offensive functions directed at people.

The genuinely new problems

Beyond any individual system, AI in warfare raises four issues that earlier technology never posed with this force.

• Distributed accountability: when an unlawful strike begins with a flawed automated recommendation, responsibility spreads across commander, operator, developer and data supplier, while the law is built around identifiable human decisions.
• Compressed time: competitive pressure rewards deciding faster than the adversary, and that pressure eats into the margin for verification and de-escalation.
• Opacity and verification: a model trained on classified data is hard to audit and hard to place under judicial or parliamentary scrutiny, which also complicates any verifiable arms-control agreement.
• A lower threshold for force: the perception of reduced risk to one's own side can make interventions politically cheaper, an effect already documented with armed drones.

None of this is fixed by better engineering. These are questions of law, doctrine and democratic oversight — a point on which the official directives governing autonomy and the campaigners seeking to ban it largely agree.