Frequently asked questions: Artificial Intelligence (AI) in the military domain

Autonomous weapons systems (AWS) are weapon systems that, once activated, can select and apply force to targets without further human intervention. 

“Without human intervention” means that after a human activates the system, it uses information from its environment – gathered through sensors that, for example, detect heat, light, movement, shape, velocity or other signals – to match potential targets against a generalized “target profile.” Examples of such profiles include the shape, infrared or radar signature or speed of a military vehicle.

Not all AWS incorporate AI. AWS can range from simple to highly complex systems and may or may not rely on artificial intelligence (AI), including machine learning (a type of AI system in which the system learns patterns from data and uses those patterns to carry out a particular task). Air-defence systems can be AWS, but they generally use simple, rule-based software rather than AI. 

The ICRC is concerned by an apparent trend to incorporate increasingly complex AI into AWS, which can exacerbate concerns regarding the unpredictability of effects. Other weapon systems may incorporate AI but not be AWS: e.g. the software might control functions such as surveillance or navigation. An AI-Decision Support System might provide input to feed a weapon system, but it will not be an AWS where there is a human decision-maker.

Autonomous weapons can select and apply force to targets without human intervention, making it difficult for users to predict and control outcomes. This unpredictability increases the risk of indiscriminate or otherwise unlawful attacks under international humanitarian law (IHL). AWS struggle to distinguish between civilians and combatants, recognize wounded or surrendering soldiers, or interpret complex situations. 

The ICRC is particularly concerned about AWS being deployed in complex environments, especially in populated areas where the difficulty of distinguishing military objectives from civilians and civilian objects is significantly increased. 

Yes. Certain autonomous weapon systems are already in use by armed forces. Examples include anti-missile defence systems on warships, which autonomously detect and engage incoming missiles. These systems generally operate in controlled environments and target clearly defined military objects. 

There is a clear trend towards greater autonomy in weapon systems, and an apparent military interest in loosening constraints on where, or against what, to use such weapons. The ICRC is particularly concerned about the prospect of AWS that incorporate AI in ways that make their functioning and effects unpredictable, or those which are designed to target humans directly. Such systems pose unacceptable risks to civilians, as well as to combatants who are wounded, sick, shipwrecked or who surrender.

AWS can be relatively inexpensive and easy to develop. Current technology enables their construction with few barriers, increasing the risk of proliferation.

Anti-personnel AWS are weapons that can select and attack humans directly. 

The ICRC considers these systems particularly concerning because the signals that indicate a person’s status in armed conflict – e.g., whether they are surrendering, injured or directly participating in hostilities – are highly contextual, making recognition by a machine process extremely challenging. Additionally, this status can change rapidly, and the consequences of getting it wrong will generally be tragic.

It is therefore difficult to envisage realistic combat situations where the use of AWS against people would not pose a significant risk of IHL violations.

There are also profound ethical concerns about, in effect, delegating life-and-death decisions to machines. Only humans can recognize and understand the intrinsic value of human life.

The process by which an anti-personnel AWS would function would reduce humans to data points. This undermines the fundamental principle that people should never be reduced to objects.

Experience with other anti-personnel weapons (e.g., anti-personnel landmines) shows that even if such weapons can in theory be used in compliance with IHL, in reality they overwhelmingly are not. Outright prohibition is therefore the most practical and effective solution, providing clarity and reliability for militaries and necessary protection for combatants and civilians alike.

Unpredictable AWS are those systems that do not allow a human user to understand, predict and explain the system functioning and effects. This will be the case, for example, with AWS that incorporate AI in such a way as to make their functioning opaque (the “black box” challenge). 

This effectively results in a lack of control over the weapon’s effects, rendering it indiscriminate by nature.

The push towards greater autonomy in weapon systems, and an expansion of the operating parameters of these systems is clear. Without timely international action to set clear legal and ethical limits, there is a risk that unacceptable autonomous weapons will be developed and proliferate. Once such weapons are widely deployed, the risks to civilians and global security will be far harder to contain.

The 7th Convention on Certain Conventional Weapons (CCW) Review Conference, scheduled from 16-20 November 2026, offers a critical opportunity for states to launch negotiations on a new legally binding instrument to regulate autonomous weapon systems.

On autonomous weapon systems, the ICRC calls on states to negotiate a legally binding international instrument to: 

• Prohibit unpredictable autonomous weapons: Autonomous weapons whose effects cannot be adequately predicted or controlled, which is more likely to be the case when systems incorporate non-deterministic forms of AI.
  • Prohibit anti-personnel autonomous weapons: Autonomous weapons designed or used to target humans directly.
    • Restrict the development and use of all other autonomous weapons: Limiting the types of targets, and the geographic scope, duration, situations and scale of their use.

In 2024, the ICRC submitted its views for consideration by the United Nations Secretary-General on ways to address the related challenges and concerns that autonomous weapon systems raise from humanitarian, legal, security, technological and ethical perspectives and on the role of humans in the use of force: ICRC_submission_on_autonomous_weapons_to_UNSG

These systems analyse large datasets to support military decisions, but they carry risks based on the way in which AI functions, and the lack of quality and context-specific data in armed conflicts. 

Their use also raises questions of automation bias – the tendency of humans to over-trust machine-generated recommendations or outputs, particularly in high-pressure situations or under time constraints. This may lead users to rubber-stamp AI outputs rather than apply independent judgment.

Additionally, in adversarial environments, opponents may deliberately manipulate data or exploit system vulnerabilities. 

These systems can amplify the speed and scale of errors and risk unintended escalation, especially if users fail to exercise independent judgment.

In certain circumstances, the careful and responsible use of AI DSS may facilitate quicker and more comprehensive information analysis, which can support decisions in a way that enhances IHL compliance and minimizes risks for civilians. For instance, the use of AI DSS can assist in collecting and synthesizing information from open-source repositories online about the presence of civilians and civilian objects or support weaponeering by recommending means and methods of attack that can best avoid, or at least minimize, incidental civilian harm. The effectiveness of any such tool will still be subject to inherent limitations of the technology and will depend upon access to good quality data. 

Ultimately, the impact on civilians of the use of these technologies will depend on the way they are designed and used.

Importantly, the use of AI-decision-support systems can never ameliorate targeting methodologies and other policies that do not otherwise comply with IHL; applying AI-decision-support systems within such frameworks will serve only to replicate and likely exacerbate unlawful or otherwise harmful effects faster and on a larger scale.

The ICRC has made 16 preliminary recommendations to address the risks associated with AI decision-support systems in armed conflict. These recommendations emphasize the importance of preserving human judgment and control. Key measures include:

• ensuring rigorous testing, evaluation, verification and validation, and legal reviews; 
• using reliable and high quality data; 
• mitigating risks of bias and discriminatory outcomes; 
• ensuring meaningful human engagement and the ability to challenge AI outputs; 
• strengthening user training, including on automation bias; 
• conducting after-action reviews. 

The ICRC also emphasizes that certain uses of AI must be prohibited, such as in nuclear command and control systems or in autonomous weapon systems that target humans directly or produce unpredictable effects.

The recommendations are guided by a human-centred approach aimed at ensuring that AI systems support human decision-making, rather than impair or replace it.

“Cyber operations” are operations against a computer, a system/network or another connected device, through a data stream, when used as means or methods of warfare. In short, using a computer to damage another computer system, network or data.

Cyber operations may disrupt, disable or physically damage civilian services, infrastructure and data. They risk resulting in injury or death of people, for instance in hospitals when medical devices are disabled, data is deleted or electricity is cut through cyber operations.

IHL imposes limits on cyber operations during armed conflicts. ICRC works bilaterally and multilaterally with states and hacker groups to ensure that existing IHL rules are interpreted and applied in a manner that ensures adequate protection for civilians and civilian infrastructure and data, including medical facilities and humanitarian organizations. However, if existing rules of IHL are interpreted in ways that undermine the protective function of IHL in the ICT environment, additional rules would need to be developed to strengthen the existing legal framework. 

Under IHL, technology companies and their employees are, prima facie, civilian objects and are protected against attack. Such protection can, however, be lost if the company’s personnel carry out certain activities or if the company’s property is used in certain ways. For instance, where specific tech company infrastructure or services are used by a warring party to make an effective contribution to military action and its total or partial destruction, capture or neutralization, in the circumstances ruling at the time, offers a definite military advantage to the adversary, it becomes a military objective.  

The ICRC therefore encourages technology companies to carefully assess how their products and services might be used in armed conflict. Companies should also take steps to avoid contributing to violations of IHL and to mitigate the risks their technologies may pose to civilians.

Yes, the ICRC actively engages with technology companies and other stakeholders involved in the development of AI systems. These discussions focus on familiarising these companies with IHL, including how their services and activities may expose their staff, assets and customers to harm. They also focus on how AI systems can be designed and developed to better support compliance with IHL and reduce civilian harm.