Introduction 

Militaries are rapidly integrating artificial intelligence (AI) into the making of life-and-death decisions: whom to target, what force to use, and how to weigh the expected harm to civilians. The adoption of AI in the military domain is now a global phenomenon, evident in the conduct of parties to armed conflict, in the priorities of major military powers, and in the investments of many other states.[1]

Human Rights Watch has conducted research and advocacy for more than a decade on the human rights and humanitarian consequences of removing meaningful human control from decisions about the use of force. Human Rights Watch has also long investigated the impact of AI technologies on human rights across a wide range of sectors, including public sector services, law enforcement, migration, and employment.[2] In addition to our extensive research on the human rights implications of autonomous weapons systems,[3] we have also documented Israel’s use of automated digital tools in the military domain – some of which use AI – in the Gaza Strip and assessed their risks to civilian protection and alignment with international human rights law and standards.[4]

This briefing, prepared by Human Rights Watch for the informal multi-stakeholder exchanges convened in Geneva in June 2026 pursuant to United Nations General Assembly resolution 80/58, sets out what states should do, through a series of recommendations to promote compliance with international humanitarian and human rights law and protect civilians in response to the threats posed by AI in the military domain. After providing technical and policy context, the briefing highlights three concerns. 

First, the adoption of AI-enabled capabilities is outpacing the testing, evaluation, verification, and validation (TEVV) on which the lawful use of any new means or method of warfare depends. In practice, the battlefield has become a testing ground. Moreover, AI systems resist traditional testing and evaluation approaches. They can only be partially assessed through formal methods and simulations, with their opacity further limiting reliable evaluation. A system fielded with insufficient understanding of its behavior precludes compliance with international humanitarian law and international human rights law. 

Second, AI is being used to plan and conduct attacks at a speed and scale that can make meaningful precautions difficult or impossible. A party’s deliberate compression of the time available to verify targets and assess civilian harm cannot excuse a failure to take the precautions international humanitarian law requires. 

Third, AI used for decision support can degrade rather than augment the human judgment on which legal compliance depends. Automation bias, the opacity of machine learning systems, and the substitution of statistical inference for the qualitative judgments that international humanitarian law demands all erode a decision-maker’s capacity to apply the law. This is especially concerning because AI systems can produce false or misleading information and can reproduce and amplify existing patterns of discrimination, often in ways that are difficult to detect.

The danger is that the standard of civilian protection drifts downward as states accept reduced human judgment and treat probabilistic outputs as substitutes for legal assessment. 

The June 2026 UN exchanges are an opportunity for states to set the agenda for how the most urgent humanitarian and human rights concerns raised by military AI are addressed—both by acknowledging and clarifying how existing law applies and by identifying where new law may be required. Human Rights Watch urges states to seize this moment. 

Recommendations

Human Rights Watch joins Stop Killer Robots, a global campaign of more than 300 civil society organizations that we co-founded, in its call for action in light of these issues. Until rules, standards, and guarantees are put in place to address the grave concerns raised by military AI, we call on states to:

• Impose an immediate moratorium on using any AI system, including generative AI, as the basis for targeting decisions.
• Provide immediate transparency on the role of AI in militaries’ use of force, and its impacts on civilians and their rights. 

The moratorium should remain in place until the following necessary steps are taken to:

• Ensure that systems that contribute to the use of force, including targeting, do not use data gathered in ways that infringe on human rights and other bodies of international law, as well as unreliable or otherwise inappropriate data. 
• Ensure that all systems that are used in military operations are operated with meaningful human control, including with regard to speed and scale, which means ensuring that they are within the cognitive capacities of users.
• Ensure that life-and-death decisions are not de facto automated, and ensure that there is always deliberative and informed decision-making in the use of force.
• Require predictability, reliability, explainability, and traceability in all systems, as well as transparency about their uses and impacts.

Complementing these calls, Human Rights Watch further recommends that states:

• Reaffirm that the use of AI should not lower the standard of civilian protection. A central risk in the adoption of AI in military use of force is that the standard of civilian protection drifts downwards through practices that accept limited human judgment and substitute probabilistic outputs for qualitative judgments required for compliance with international law. At the national level, states should articulate in their own doctrine how the standard of civilian protection will be maintained if AI is used in the use of force. 
• Acknowledge that the adoption of AI systems, including machine learning, algorithmic, and automated systems, into military pipelines introduces new and novel challenges regarding transparency and accountability for violations, due to the technical design of the systems, the way that they are licensed and contracted, and the role of private sector actors in development and deployment of military AI.
• Establish a process to develop common good practices on testing, evaluation, verification, and validation (TEVV), risk forecasting and harm mitigation, legal reviews, auditing and reporting, and post-action review.
• Focus future discussions on the application of international law to the role of AI in the use of force and the development of steps to address the range of humanitarian concerns raised by military AI. Topics could include red lines on the use of AI in conflict and the application of precautions in attack to the military use of AI.
• Prioritize the discussion of AI applications and uses that are of the greatest humanitarian concern. The phrase “AI in the military domain” encompasses a heterogenous set of applications that raise distinct legal and humanitarian concerns and that call for distinct responses. 
• Establish accessible mechanisms for the adequate fulfillment of rights to remedy and redress, with specific regard to AI systems in the military domain.

Context

How Militaries Use AI

AI is an umbrella term used to describe a suite of general-purpose technologies. Its effects materialize through integration or interactions with data and other digital systems, with data input from sources including sensors, autonomous platforms (e.g. navigation or predictive maintenance), munitions, cloud platforms, and data obtained through surveillance. Treating AI as a singular “weapon,” to be defined and regulated in isolation, misunderstands the nature of AI. AI is generally used to describe a variety of networked, layered, and at times interdependent systems.[5]

AI is being adopted rapidly by militaries for a range of use cases. These capabilities and potential uses can largely be grouped under five headings:

1. Reasoning (networking dispersed data points on command);
2. Learning (modifying – or “optimizing” – operation based on experience rather than explicit programming);
3. Planning (design and execution of programmed strategies);
4. Perception (mapping the environment through data accumulated via sensors); and 
5. Communications (generating information in text or audio form, for human-machine and machine-machine communication).[6] 

Examples include the use of computer vision for target identification (perception and reasoning), the use of reinforcement learning for mission and route planning of uncrewed systems (learning and planning), the integration of large language models (LLMs) into command and intelligence platforms to summarize and query multiple sources of intelligence data and to generate courses of action (reasoning and communication).[7]

These capabilities have applications in military uses of force that can be grouped under four further headings:[8]

1. Autonomy, which describes the use of AI to automate the performance of labor (such as piloting an aircraft) or cognition (such as selecting a target). Autonomous weapons that integrate AI in critical functions are a category in this heading, such as Ukraine’s Saker Scout, Russia’s Zala Lancet, Israel’s Harpy or Skystriker loitering munitions or, reportedly, the STM Kargu-2 used by the Government of National Accord Affiliated Forces (GNA-AF) in Libya.[9]
2. Insight, meaning the use of AI to generate inferences from data, including to classify persons or objects, such as Israel’s Lavender and Gospel systems.[10]
3. Decision-making, which captures the use of AI in recommending or selecting courses of action, such as target recommendation systems like Hadean’s dominAI, integrated with Palantir’s Maven Smart System, or Israel’s “Fire Factory.”[11]
4. Management, namely the use of AI for coordination, such as drone swarm coordination or the orchestration of multi-domain operations. Both China and Russia have publicly demonstrated Unmanned Aircraft (UA) swarming concepts, and the European Union is developing a border protection tool (“Roborder”) that will operate drones in swarms.[12]

Applications of Greatest Concern 

“AI in the military domain” is a broad category that encompasses the development and use of AI-enabled applications ranging from logistics, training, and predictive maintenance to command and control to autonomous weapons systems. Within that broad category, this briefing focuses on the applications that have the most direct and documented implications for civilians and their human rights. They are: AI used to support decisions on the use of force, which involves the identification, selection, and engagement of targets, and the assessment of expected civilian harm.[13] The briefing does not cover autonomous weapons systems, which both select and engage targets without human intervention – Human Rights Watch and others have separately called for a legally binding instrument with prohibitions and regulations on those specific systems.[14]

This focus is warranted given current uses of AI in the military domain. In Gaza, Human Rights Watch has documented Israel’s use of four digital tools that rely on faulty data and inexact approximations, increasing the risk of civilian harm and creating conditions in which Israeli forces may violate international humanitarian and human rights law.[15] In Russia’s war against Ukraine, the integration of AI-enabled target recognition systems and weapon guidance systems is shaping the pattern of the conflict by, for example, lowering barriers to the use of lethal force and changing how parties attack and defend vulnerabilities.[16] The US has described how it is using the technology to identify potential targets and accelerate decision-making as part of its operations in Iran.[17] Other publicly described applications include multi-source intelligence and command platforms used by NATO members; AI components in air and missile defense; and AI used in the planning and conduct of cyber operations. The diversity of applications is itself part of the policy challenge: each of these uses raises distinct legal and humanitarian concerns and each use case calls for distinct policy responses. 

Accelerated Adoption

Many states now treat military AI adoption as a strategic and operational imperative.[18] Adoption is also being shaped by close working relationships between defense ministries and suppliers of technology that uses AI, including companies whose products were not originally designed for military use. These suppliers echo and amplify the narratives emphasizing urgency and strategic and operational imperatives.[19] 

Policy Landscape

The diplomatic and policy landscape has expanded markedly in the last five years in response to the uptick in AI deployment in military campaigns. States have convened and adopted commitments in relation to military AI, such as the Political Declaration on Responsible Military Use of Artificial Intelligence and Autonomy, the AI Action Summit’s Paris Declaration on Maintaining Human Control in AI Enabled Weapon Systems, and outcome documents of successive Responsible Artificial Intelligence in the Military Domain (REAIM) Summits.[20]

These initiatives have raised awareness and surfaced areas of shared concern. But they are also limited to high-level and abstract language of “responsibility.” This is imprecise framing that risks erasing the obligations that states already owe under international law and obscuring where the development of new rules is needed.[21]

The dual-use nature of some of these technologies also presents a complex regulatory question regarding the control of AI systems used in civilian (state and commercial) as well as military contexts, applicable across software and hardware.[22]

To translate growing political and policy interest in AI in the military domain into protection for civilians and their rights, states need to be specific about the existing obligations that apply, how they apply, what compliance requires in practical terms, and where further steps are required to address risks that existing rules do not adequately regulate.

The June 2026 informal exchanges present an opportunity for states to set the agenda for how the most urgent humanitarian and human rights concerns associated with militaries’ use of AI are addressed—both by clarifying the application of existing international law and by identifying where additional measures are needed, including a moratorium and red lines on the use of AI in military contexts, alongside commitments to transparency, accountability, and remedial mechanisms.[23] States should use these exchanges to prioritize such concerns in the framing and agenda-setting of subsequent international discussions.

Analysis and Findings

AI Adoption and Use is Outpacing Testing Required for Lawful Use

States are rushing to purchase, adopt, and scale systems that are marketed as using AI to offer indispensable strategic and operational advantage. This pressure is resulting in military AI capabilities being put into use before their performance can be reliably and independently tested and evaluated to a universally established standard.[24] This raises a concern that a party to an armed conflict will lack the information needed to comply with its international humanitarian law obligations, particularly information that is needed to anticipate and control the effects of weapons and means of warfare that utilize AI. 

International humanitarian law requires those who plan and decide upon attacks to anticipate the effects of the means and methods of warfare they employ.[25] The obligations to take constant care to spare the civilian population, to do everything feasible to verify that targets are military objectives, and to assess and, where necessary, limit expected incidental civilian harm all presuppose that the attacker can foresee, with sufficient confidence, how the capability in question will perform in the circumstances of its use.[26] The prohibition on indiscriminate attacks makes the same assumption: it extends to any attack employing a means or method of warfare the effects of which cannot be limited as required by international humanitarian law.[27] A capability whose effects cannot be anticipated cannot be limited, and its use against persons or objects cannot satisfy these rules, whether or not civilian harm in fact eventuates in a particular instance. The obligation to review new weapons, means, and methods of warfare is one important mechanism for surfacing this information.[28]

The ability to anticipate effects depends on rigorous testing, evaluation, validation, and verification (TEVV) capable of establishing how a system behaves and how its behavior may change over time and in different environments of use. AI capabilities challenge traditional TEVV methodologies. Formal methods to verify the reliability of hardware and systems are challenging to apply to systems that use AI, especially probabilistic systems. Reliability is therefore primarily evaluated through simulations and operational testing. Such empirical tests and evaluations can only provide a partial picture of how systems perform and potentially fail. The opacity of systems-based automation complicates the equation. In the absence of an explanation for how a system reaches its output from a given input, it can be difficult for engineers, operators, and auditors to identify where failures might originate.[29]

Where TEVV cannot keep pace with adoption, neither legal reviews nor operational decisions rest on adequate information. The same commercial incentives and operational tempo that produce premature fielding may operate in other contexts in a way that puts civilians (and combatants) and civilian objects at risk. Withdrawing a prematurely fielded system from service is not a sufficient safeguard. To direct an attack through a means whose effects the attacker cannot anticipate is to act without the information that distinction, proportionality, and precautions require. Responsible iteration is welcome, but it cannot be conducted on the battlefield, against real persons and objects, as a substitute for establishing a system's behavior before it is fielded.[30] Furthermore, data gathered from one conflict may be an inappropriate source of testing information for another.[31] 

Use of AI Facilitates Attacks at Scale and Speed that Interfere with Meaningful Precautions

Militaries seek to convert observation into action—including using force against targets—as quickly as possible, and the volume of sensor-derived observations available to them is growing rapidly. AI-enabled systems are valued precisely because they process that data and generate outputs, such as target nominations and recommended courses of action, at scale and at speed. The use of AI in identifying and selecting targets and in assessing expected civilian harm and anticipated military advantage facilitates the use of force at a speed and scale that can make the discharge of obligations of precautions in attack difficult or impossible.[32] 

Whether civilians are mistaken for targets or harmed in attacks directed at military objectives, the consequence is harm—death, injury, and the destruction of civilian objects. A reliance on AI without adequate safeguards has the potential to increase the scale of harm dramatically.

Recent developments exemplify the reduced speed and greater scale of AI-enabled systems. For example, British forces trialed Project ASGARD, a digital targeting system that is still at the prototype stage, during a NATO exercise in Estonia in 2025, where it was presented as being able to collapse the targeting cycle from hours or days to minutes as a means of doubling the army’s lethality.[33] During Operation Epic Fury—the US-led air and missile campaign against Iran that began on February 28, 2026—US forces reportedly relied on Palantir’s AI-enabled Maven Smart System integrated with Anthropic’s Claude LLM to identify priority targets and help select the munitions used against them.[34] Senior officials attribute the campaign’s tempo and widespread scale in substantial part to AI-enabled targeting.[35] The US Defense Department’s own usage figures record a surge in reliance on the system over the same period.[36] Some critics attributed targeting errors and civilian casualties to over-reliance on AI.[37]

Parties to a conflict must take constant care to spare the civilian population, and must take all feasible precautions in attack—including in the choice of means and methods of warfare, in verifying that the objective is a military one, and in assessing the expected incidental harm to civilians.[38] What is “feasible” is contextual; it takes account of the circumstances ruling at the time, including operational constraints.[39] But it is not infinitely elastic, and a party should not be able to manufacture the very time pressure that it then invokes to excuse the precautions it failed to take. Where a party’s own choice of an AI-enabled method of warfare compresses the time available for verification and harm prevention and reduction, that self-imposed compression cannot be used to excuse a failure to take feasible precautions.[40] 

Use of AI Degrades Rather than Augments Decision-Making Essential to Legal Compliance

International humanitarian law requires those who plan and conduct attacks to make qualitative context-dependent assessments—whether a person or object is a lawful target, whether expected civilian harm would be excessive in relation to the anticipated military advantage, and what precautions are feasible. This allocation of responsibility presupposes that the decision maker is genuinely able to exercise judgment: that they have the training, information, time, and procedures necessary to do so. The use of AI for insight (generating inferences from data), decision-making (recommending or selecting courses of action), and management (coordinating action across systems) can undermine each of these preconditions.[41]

Automation bias leads users to defer to machine-generated outputs, including when those outputs may be wrong. Statistical inferences are perceived with a confidence they do not warrant. The inferences an AI system draws from data may be unreliable, because the data are incomplete, unrepresentative, or contaminated, or because the inferential step itself is flawed. The difficulty is that this unreliability may be invisible at the point of use and in subsequent reports.[42] When system operators cannot perceive the unreliability of the inferences they are given, they cannot comply with the principles of distinction and proportionality: the decision-maker is, in effect, applying the law to a representation of the situation that they have no means of testing, and that may be wrong in ways the interface does not disclose.[43]

Relatedly, opacity prevents system operators from understanding how an output was generated, and therefore from interrogating it. Users cannot, or find it difficult to, understand how an AI-enabled system produced an output. Opacity may be: technical (no human-readable evidence of reasoning); data-based (no information on what went into the system); based on the unique interaction between the chain of components (sensors, classifiers, interfaces etc.); institutional (e.g. classifications, proprietary secrets); and/or procedural (lack of documentation, limited time, lack of training, no processes). When users cannot see the unreliability of the inferences, they cannot discharge their international humanitarian law obligations.[44]

When systems use AI to identify potential targets or civilian harm, users may lack or not take the time or information to make a human judgment, and in effect rely on system output (automation bias) even when the output is deficient (e.g., the output misidentifies a target as lawful when it is in fact unlawful, or understates the expected civilian harm or overstates the anticipated military advantage). The Israeli military’s reported use of the “Lavender” system in Gaza provides an example. According to investigative reporting from +972 Magazine, the system was used to generate tens of thousands of nominations of individuals suspected of affiliation with Palestinian armed groups; intelligence officers reportedly spent on the order of 20 seconds reviewing each nomination before a strike,[45] in many cases doing no more than confirming that the target was male, despite awareness that the system produced erroneous identifications in a known proportion of cases.[46] One officer reportedly described having no added value beyond serving as a rubber stamp for the system’s output.[47] 

Application of International Human Rights Law and Standards

Given that international human rights law applies during both peacetime and wartime, it covers all circumstances relevant to the development and deployment of AI systems in the military domain. States should in the context of military AI especially consider the ability of actors to meet their human rights obligations and responsibilities with regard to the rights to life, privacy, non-discrimination, expression and opinion, peaceful assembly, dignity, and remedy. Human Rights Watch has analyzed the risk that autonomous weapons systems pose to human rights; many of these concerns apply to AI systems that are not classed as autonomous weapons systems.[48] The risk of privacy violations is notable, as AI systems deployed in military use cases are often “dual use” and developed and deployed in non-military settings where widespread and indiscriminate data collection may constitute mass surveillance, which is incompatible with international human rights law.[49] Furthermore, the repurposing of data from non-military contexts poses additional problems regarding tests providing for interference with privacy rights, which requires that data collection must be necessary and proportionate and in pursuit of a specific legitimate aim.[50] Where data has been repurposed outside of a stated legitimate aim, it must not be used. 

The distinct risk that AI technologies pose to non-discrimination rights is well documented.[51] In the context of military AI, companies’ programming policies, training datasets, and auditing datasets influence a system’s design and output with discriminatory impact. Algorithmic bias can disproportionately negatively affect people with vulnerable and marginalized identities, including through their race and ethnicity, gender, disability status, and/or immigration status.

AI systems present challenges to transparency, oversight, and accountability through their opaque design, a challenge that is compounded when these systems are licensed in military and classified contexts and not subject to adequate scrutiny and regulatory safeguards. Transparency of AI systems in the military domain on both a technical and deployment level is vital for states to meet their obligations to fulfill remedy and redress rights.[52] The use of AI in the military domain therefore creates an accountability gap that significantly hampers access to adequate remedy and redress measures for rights violations and abuses.[53]

Private sector actors, including technology companies, contracting with state agencies to provide products or services for military or classified purposes have a responsibility to respect human rights. Under the UN Guiding Principles on Business and Human Rights, companies have a responsibility to avoid causing or contributing to human rights abuses, and to address risks directly linked to their business operations and relationships.[54] In conflict-affected contexts, the risk of gross human rights abuses is heightened and, therefore, due diligence by businesses should be heightened accordingly.