International Humanitarian Law & AI
International humanitarian law — the body of rules that seeks to limit the effects of armed conflict on people and property — rests on four foundational principles that have governed warfare since the Geneva Conventions and their Additional Protocols. These principles were developed for human combatants making human decisions at human speeds. Their application to systems that make algorithmic assessments in milliseconds is the central legal challenge of the AI weapons age.
The Accountability Gap
When an autonomous or AI-assisted system kills civilians in violation of IHL, the question of legal accountability becomes deeply problematic. Traditional military law places accountability on the human commander who ordered or failed to prevent an unlawful attack. In AI-assisted systems, the chain of causal responsibility diffuses across multiple nodes:
- The programmer who wrote the targeting algorithm
- The data scientists who selected and labeled training data
- The procurement officer who specified the operational requirements
- The commander who authorized deployment of the system
- The operator who approved the AI-generated targeting recommendation
- The executive who decided the system was safe enough to deploy
This diffusion of responsibility creates what legal scholars call the "accountability gap" — a situation in which no single individual can be said to bear the legal responsibility that IHL requires, even when civilians die as a direct result of the system's operation. No current international legal framework adequately addresses this gap.
The Martens Clause
"In cases not covered by this Protocol or by other international agreements, civilians and combatants remain under the protection and authority of the principles of international law derived from established custom, from the principles of humanity and from the dictates of public conscience."
— Additional Protocol I to the Geneva Conventions, 1977 (Martens Clause)
First articulated by Russian diplomat Fyodor Martens at the 1899 Hague Conference, the Martens Clause has become the primary legal argument for those who contend that autonomous weapons are prohibited even in the absence of a specific treaty ban. The argument runs as follows: public conscience — as evidenced by widespread civil society campaigns, academic consensus, and the stated positions of many states — has consistently held that allowing machines to decide to kill humans violates fundamental principles of humanity. Under the Martens Clause, this public conscience has the status of law.
The counterargument, advanced by states developing autonomous systems, is that the Martens Clause is aspirational rather than prescriptive, and that states have historically and legally deployed weapons systems whose individual targeting decisions are not made by humans — landmines, area-effect munitions, some air defense systems. The debate is unresolved and unlikely to be resolved without a specific treaty.
Convention on Certain Conventional Weapons
The Convention on Certain Conventional Weapons (CCW) has been the primary international forum for negotiations on Lethal Autonomous Weapons Systems (LAWS) since 2014. The CCW framework, which produced bans on blinding laser weapons and restrictions on landmines, was selected as the venue because it includes major military powers as parties and operates by consensus. Its progress on LAWS has been, by every honest assessment, glacially slow.
CCW holds its first informal meeting on LAWS. States agree the issue warrants attention. No further commitments.
CCW states parties agree to include LAWS as a formal agenda item. Group of Governmental Experts (GGE) established.
Multiple GGE meetings held. States articulate widely divergent positions. No binding outcomes produced.
GGE agrees on 11 non-binding guiding principles for LAWS. Affirms IHL applicability, human responsibility, need for precaution. No enforcement mechanism.
CCW meetings suspended or curtailed. Momentum stalls. Technology continues advancing without governance progress.
Sessions resume but fundamental disagreement on whether to pursue a ban or regulatory framework prevents progress. US, Russia, India oppose ban treaty.
UNGA passes non-binding resolution urging states to take steps to address risks of LAWS. 164 votes in favor. US, Russia, India abstain.
GGE chair sets 2026 as target for substantive framework agreement. Major powers privately signal unwillingness to accept binding constraints.
The year in which a binding framework was supposed to be achieved. The Iran conflict confirms AI is already directing targeting in live operations. The deadline will almost certainly be missed.
State Positions
| State / Group | Position | Reasoning |
|---|---|---|
| Austria, New Zealand, Costa Rica, 70+ states | Pre-emptive Ban | Autonomous kill decisions inherently violate human dignity; IHL cannot be satisfied without meaningful human control. |
| European Union | Binding Regulation | Meaningful human control standard required in binding treaty; ban may be too broad given dual-use technology. |
| United Kingdom, France, Germany | Political Declaration | Prefer non-binding political commitment with national implementation; skeptical of verification feasibility. |
| United States | Opposes Binding Treaty | Existing IHL sufficient; new treaty premature; human-machine teaming requires operational flexibility. Advocates for "responsible" development guidelines. |
| Russia | Opposes Any Ban | Autonomous systems are legitimate military tools; sovereignty concerns; verification impossible. Has vetoed multiple procedural advances. |
| China | Nuanced Position | Supports ban on fully autonomous weapons with no human control; opposes constraining human-machine teaming or intelligence fusion systems. |
| Israel | Opposes Ban | Self-defense and operational security requirements demand autonomous capability; cites persistent threat environment. |
| India | Neutral to Opposed | Developing indigenous AI weapons capability; caution about premature binding constraints on technology still evolving. |
The CCW operates by consensus, meaning any state can block any outcome. Russia has used this power systematically to prevent procedural advances. The US has avoided formal opposition while blocking substantive progress through definitional disputes: if LAWS cannot be defined precisely, a treaty cannot be written. The major AI weapons developers have every strategic incentive to delay binding frameworks while their capabilities mature. The states most harmed by autonomous weapons — those without the resources to develop counter-capabilities — have the votes but not the leverage.
Notable Frameworks & Policies
In the absence of binding international law, a patchwork of national directives, alliance principles, and civil society frameworks has emerged to govern — or at least describe — how AI weapons should be developed and used. None are enforceable across borders. None have prevented the deployment of AI targeting systems in live combat. But they represent the current state of institutional thinking and provide the basis for future accountability claims.
For the most current state of US defense AI policy and military doctrine, see our dedicated coverage sections. For country-specific AI weapons programs and positions, see our country profiles.
The Human-in-the-Loop Spectrum
The debate over autonomous weapons frequently hinges on where human judgment sits in the targeting process. Three conceptual positions on this spectrum have emerged as the dominant framework for policy discussion. The distinctions are important but unstable — the operational tempo of AI-assisted warfare continuously erodes the meaningful content of categories that sound robust in policy documents.
Nations: All nations claim to operate here. Formal doctrine for all CCW states parties. Reality increasingly diverges from doctrine as tempo increases.
Status: The legal minimum most IHL advocates argue is required. Becoming structurally impractical at scale under AI-assisted operational tempo.
Nations: US, Israel, UK, South Korea (Samsung SGR-A1). Most advanced militaries in time-critical defensive scenarios.
Status: The de facto operational standard for AI-assisted targeting under tempo pressure. The human cannot meaningfully review AI-generated packages at the speed they arrive.
Nations: No state officially operates here. Credible evidence suggests Turkey, Russia, and potentially Israel have deployed systems that functionally operate at this level.
Status: Officially prohibited by most national policies. Technically and operationally active in specific scenarios. The future trajectory of all AI weapons programs.
The three categories describe architectures, not realities. A system designed as "human-in-the-loop" in which the human reviews 400 AI-generated targeting packages per shift, approving 390 of them without independent verification, is functionally operating at human-on-the-loop or below. The meaningful content of human control is not measured by whether a human button-press precedes a strike. It is measured by whether the human possessed the information, time, and cognitive capacity to exercise genuine judgment. By that standard, the entire taxonomy has been quietly hollowed out by operational tempo.
For technical details on specific autonomous weapons systems and their control architectures, see our systems database and threats analysis.
Case Studies in AI Ethics Failures
The theoretical concerns about AI weapons — the accountability gap, algorithmic bias, the erosion of human judgment — have materialized in documented real-world failures. The following cases represent the most consequential and best-documented instances where AI-assisted or autonomous systems produced outcomes that violated or structurally undermined IHL principles.
Reporting by +972 Magazine and Local Call, based on testimony from Israeli intelligence officers, revealed the existence and operation of an AI system called Lavender, used extensively in the Gaza conflict beginning in October 2023. Lavender was trained to identify suspected Hamas operatives based on behavioral, communications, and association patterns. The system generated a database of individuals it assessed as likely militants — ultimately more than 37,000 people.
Officers testified that the system was used as a "kill list" with minimal independent verification. For low-ranking targets, the military accepted a pre-defined "acceptable" civilian casualty ratio — in some periods, up to 15 or 20 civilian deaths per targeted individual — and strikes were authorized on this basis. One officer described the process as "rubber-stamping" the AI's output. Intelligence officers reported spending approximately 20 seconds per target reviewing AI-generated dossiers before authorization.
A companion system, called "Where's Daddy," identified when a target entered their family home, enabling strikes at times of maximum civilian presence under the logic that collateral damage was pre-authorized within the accepted ratio. A third system, "The Gospel," managed target generation for infrastructure. Together, they constituted an industrial-scale AI targeting pipeline operating in a densely populated urban environment.
A 2021 UN Panel of Experts report on Libya described what may be the first documented lethal engagement by a fully autonomous weapons system operating without human authorization. Turkish-manufactured STM KARGU-2 loitering munitions, deployed by forces allied with the Government of National Accord, were reported to have "hunted down and remotely engaged" retreating forces from the Libyan National Army. The panel's language was careful but the implication was significant: the system was operating in an autonomous mode without a human directing individual engagements.
STM, the Turkish manufacturer, disputes this characterization, maintaining that the KARGU-2 requires human authorization for engagement. The ambiguity is itself revealing: a system designed with an autonomous mode, deployed in a chaotic combat environment, operating in conditions where the distinction between authorized and autonomous engagement cannot be externally verified. The incident represents either the first confirmed autonomous kill in warfare, or a system whose design and deployment parameters make confirmation impossible. Both possibilities are alarming.
The United States drone strike program — conducted under authority flowing from the 2001 Authorization for Use of Military Force, and spanning multiple administrations and multiple countries — provides the closest historical parallel to AI-assisted targeting at scale. The program relied on signature strikes: engagement authorizations based on pattern-of-life analysis rather than positive individual identification. A person could be designated as a target not because they were identified as a specific individual but because their behavioral patterns matched a threat model.
Investigations by the Bureau of Investigative Journalism, Airwars, and multiple academic researchers found that in specific periods and geographies, civilian casualty rates in signature strikes were dramatically higher than official characterizations suggested. Bureau of Investigative Journalism data found that in Pakistan between 2004 and 2018, between 910 and 2,200 civilians died in US drone strikes. In Yemen, the ratio of civilian to militant deaths in some strike categories reached 90 percent or higher in specific documented cases.
The structural parallel to AI targeting is precise: pattern-of-life analysis, algorithmic threat assessment, compressed decision timelines, and a systematic undercount of civilian casualties driven by definitional choices. The signature strike doctrine was the human precursor to the AI targeting systems now being deployed. Its casualty record provides empirical context for what AI-accelerated versions of the same logic produce.
The problems documented in civilian predictive policing systems — where algorithmic risk scores trained on historically biased data systematically over-flag individuals from marginalized communities — translate directly into military targeting contexts. In both domains, an AI system is trained on historical human decisions that embedded human biases, then deployed to make new decisions that inherit and amplify those biases at scale.
In targeting applications, the training data for systems like Lavender consists of intelligence about known or suspected combatants. If that intelligence systematically over-represents certain communities, professions, or communication behaviors as threat indicators — because they share characteristics with previously targeted individuals rather than because they represent actual threat — the AI will learn to flag those characteristics. The system does not know that the pattern in its data reflects historical targeting bias rather than actual threat. It learns what it was taught.
In Gaza, reporting indicated that Lavender flagged individuals for patterns including use of certain communication applications, association with known Hamas members (broadly defined), and residential proximity to previously targeted locations. Multiple of these indicators are shared by large proportions of the civilian population in a densely networked, politically complex urban environment. The system's 90 percent confidence threshold still implies a 10 percent error rate — on a list of 37,000, that is 3,700 people incorrectly identified as militant targets.
For detailed incident reporting, see our incidents database. For country-specific case studies, see case studies.
The Arguments
The debate over autonomous weapons is not simply technical. It is a collision of competing values: precision versus accountability, military necessity versus civilian protection, strategic advantage versus humanitarian principle. Below are the strongest versions of the arguments, including positions held by serious people with considered views.
Key Voices & Organizations
The debate over AI weapons is shaped by a relatively small number of institutions and individuals whose work has defined the terms of the conversation. Below are the most consequential voices — the organizations that litigate the issue internationally, the researchers who supply the technical and philosophical arguments, and the writers who have made the stakes comprehensible to policymakers and the public.
For ongoing policy developments, see our policy tracker. For country-specific programs and positions, see our country profiles. For the latest incidents involving autonomous systems, see our incidents database. For the doctrine frameworks governing these systems, see our doctrine section.