Frequently asked questions: Autonomous weapon systems

Autonomous weapons systems (AWS), also commonly referred to as “killer robots” or “slaughterbots”, are weapons that, once activated, can select and apply force to targets without further human intervention.

“Without human intervention” means that once a person activates the weapon, it works on its own from there. It uses sensors to detect phenomena such as heat, light, movement, shape, or speed, and uses this input to match potential targets against a general "target profile." 

For example, a profile might include the shape, speed, or heat/radar signature of a military vehicle.  If the weapon assesses that the detected object matches these predefined characteristics, it can self-initiate a strike without a person selecting that specific target at the time of attack.

Yes, autonomous weapon systems are already a reality.  Examples include:  

• Air and missile defence systems, including close-in weapon systems on warships and land-based systems protecting bases, which autonomously detect, track and intercept incoming missiles, rockets or mortars – faster than a human could.
• Active protection systems on tanks, which automatically detect and engage incoming projectiles or missiles before they hit the vehicle.
• Some sensor-fused munitions, which release submunitions with sensors that search for and strike target types such as armoured vehicles, without a human selecting each individual target.
• Some loitering munitions, which can reportedly search an area and engage targets, such as radars, that match a pre-programmed profile. 

Currently, these systems are largely ‘rule-based’ or ‘deterministic’ systems, and generally operate in controlled environments against a narrow range of military objectives such as enemy missiles or other projectiles, military radars or warships. 

But the ICRC is concerned about an apparent trend towards incorporating increasingly complex artificial intelligence (AI) into these autonomous weapons systems, and an apparent military interest in loosening constraints on where, or against what, to use such weapons, which can exacerbate the unpredictability of their effects. This would create significant risks for civilians and civilian objects in armed conflict. 

Autonomous weapon systems raise serious humanitarian, legal and ethical concerns because they can select and strike targets without further human intervention. These concerns are not linked only to the technology itself, and so cannot necessarily be addressed through engineering fixes or design choices.

1. Humanitarian risks: proliferation and conflict escalation. Like any other weapon, autonomous weapons will proliferate, and be used in violation of international humanitarian law (IHL), including by non-state armed groups and criminal organizations. Indeed, they are particularly vulnerable to proliferation, as the sensors, software and other commercially available components become cheaper and more accessible. There is also a risk that the use of autonomous weapons will drive conflict escalation, as they dramatically accelerate the process of selecting and engaging targets, heightening the risk of miscalculations and reducing the space for human intervention and de-escalation. 

    

2. Legal challenges: difficulty ensuring compliance with international humanitarian law and human rights law. Because autonomous weapons select and strike targets without human intervention, it makes it hard for human operators to predict the circumstances of an attack in advance, and assess whether the attack will comply with IHL. This is particularly challenging as autonomous weapons become more complex (and consequently less predictable) and where they are deployed in populated or changing environments. 

   This increases the risk of attacks that violate IHL’s rules on distinction, proportionality and precautions in attack. These rules require context-dependent assessments, in the circumstances ruling at the time; they cannot simply be ‘coded in’ in advance, or reduced to technical indicators. Determining whether a person is a civilian directly participating in hostilities, whether a combatant is surrendering or otherwise protected, or whether expected civilian harm would be excessive requires qualitative understanding of factors such as intent, conduct and context that a machine calculation cannot provide. 

   Unpredictable or opaque systems compound these challenges: if users cannot adequately understand, predict and explain how a weapon will behave and what effects it may produce, assessing the lawfulness of its use will not be possible. Such systems may also complicate questions of responsibility and accountability when unlawful harm occurs.

    

3. Ethical dilemma: effectively delegating life-and-death outcomes to machines. Beyond the humanitarian and legal risks lies a fundamental ethical question: should a machine ever be allowed to determine whether a human being lives or dies? Only humans can recognize the intrinsic value of human life and exercise moral judgment and restraint. When a machine targets a person on the basis of sensor data and a target profile, that person risks being reduced to a set of data points, removing human judgment from the decision to apply lethal force. This concern does not disappear simply because technology becomes more accurate. For the ICRC, retaining meaningful human control over decisions involving life and death is a moral imperative.

International humanitarian law already applies to autonomous weapons. But the ICRC believes existing rules do not adequately address all of the specific humanitarian, legal and ethical risks they raise. This is why we are calling for new, legally binding international rules to prohibit unacceptable autonomous weapons and place strict limits on all others.

Autonomous weapons are not necessarily unlawful. Existing IHL applies to the use of autonomous weapons, as it does for any other weapon. 

IHL rules prohibit certain types of weapons, such as ones that are inherently indiscriminate. This is particularly relevant for an autonomous weapon that functions in an opaque way (the “black-box” challenge), which may prevent humans from being able to understand, predict or explain the system’s output. This prevents a user from being able to direct it against a specific military objective and effectively results in a lack of control over the weapon’s effects, making it indiscriminate by nature.

Any attack with any weapon also needs to comply with IHL’s rules on the conduct of hostilities: it must only be directed at a specific military objective, it cannot be expected to cause excessive incidental civilian harm, and all feasible precautions must be taken to avoid civilian harm. But with autonomous weapons, these things can be difficult to assess and ensure in advance, since the user does not choose, or even know, the specific person or object that will trigger the AWS to strike, or precisely when and/or where the strike will occur. These details will depend on the machine processes (sensors, software, actuators) of the weapon system, and the sensory input from the operational environment. 

So how can the user then ensure that any strike by the weapon system will only be against a legitimate military target, and not a civilian or someone who is wounded or surrendering? How can they assess, in advance, that any incidental harm will not be excessive, even though the situation on the ground might shift between activation and impact? And what if it becomes apparent that the target is not a military objective: IHL says they have to do everything feasible to cancel or suspend the attack, but with an autonomous weapon this may not be possible.

Meeting these legal obligations with an autonomous weapon is likely to be possible only if certain types of autonomous weapons are ruled out altogether, and strict constraints are imposed on the operating environment for others, in order to limit the number of potential outcomes.

This is why the ICRC and many others call for new, legally binding rules that prohibit unpredictable and anti-personnel autonomous weapons, and place clear restrictions on all others.

Not all weaponized drones are autonomous weapons. 

An armed drone piloted remotely by a human operator, who identifies the target and decides when to fire, is not an autonomous weapon. But a drone can become an autonomous weapon if, once activated, it selects and applies force to a target without a human choosing the specific target. 

The key test is not whether the weapon flies autonomously, but whether it has autonomy in the critical targeting functions: selecting and applying force to targets.

Not all autonomous weapons systems incorporate artificial intelligence (AI). 

Autonomy in weapons can range from simple to highly complex systems and may or may not rely on AI - including machine learning (when the system learns patterns from data and uses those patterns to carry out a particular task). For example, air-defence systems can be autonomous, but they generally use simple, rule-based software rather than AI. 

Other weapon systems may incorporate AI but not in their functions of selecting and engaging targets. So, a weapon system that uses AI only in navigation will not necessarily be an autonomous weapon. 

Autonomous weapons are already a reality, ranging from defensive systems that automatically engage incoming missiles to some loitering munitions and sensor-fused weapons. The ICRC’s concerns lie in trends toward systems being designed to operate over broader areas, for longer periods, and against a wider range of targets, while integrating increasingly complex forms of artificial intelligence. 

Such military and technical developments bring additional unpredictability, both in terms of how these autonomous weapons will function and how they will interact with the operating environment, creating significant risks for civilians and civilian objects in armed conflict. 

While promoters of new weapons technologies often claim that such weapons have the potential to enhance compliance with international humanitarian law and minimise risks for civilians, this is not what we have seen historically or what we are seeing now.

In many of the ongoing conflicts around the world that involve the most technologically sophisticated means of warfare, we have not seen better outcomes for civilians but rather widespread and unacceptable levels of civilian harm and devastation. The qualities advertised in new technologies have led not to precision, distinction and precaution, but to the acceleration, amplification and escalation of destruction, with appalling results for civilians.

This creates an increasingly unpredictable environment, heightening the risks of harm to civilians, and underscores the need for clear prohibitions and restrictions on AWS, to prevent practices that would erode the protections international humanitarian law is meant to guarantee.

Anti-personnel autonomous weapon systems 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.

Beyond the operational and legal risks, there's a deeper question at stake: should a machine ever be allowed to determine whether a human being lives or dies? Only humans can recognise and understand the intrinsic value of human life. If a machine were to target a person based on sensor data and a target profile, that person would be reduced to data points. The space for human judgment and moral restraint would be eliminated. This is a moral redline that we must not cross, regardless of how accurate the technology becomes.

Experience with other anti-personnel weapons (e.g., anti-personnel landmines) shows that even if such weapons can theoretically 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 is particularly the case with AWS that incorporate AI in ways that make their functioning opaque (the “black box” challenge). 

This lack of predictability effectively results in a lack of control over the weapon’s effects, rendering it a weapon that is indiscriminate by nature, which is prohibited under IHL.

Outright prohibition is therefore the most practical and effective solution, providing the necessary distinction capability for the protection of civilians and compliance with the laws of war.

The trends toward greater autonomy in weapons and a loosening of operating constraints is already clear. The gap between what machines can do and what they should be allowed to do is widening. 

ICRC President Mirjana Spoljaric and UN Secretary-General António Guterres, on 25 August, renewed their urgent joint appeal to world leaders to adopt new international rules on autonomous weapons, warning that we are now dangerously close to crossing a moral red line: the autonomous targeting of humans by machines. 

As the very scientists and engineers building these systems are also sounding the alarm and urging regulation, the world cannot afford to wait until their use becomes widespread and let delays allow technology and practice outpace regulation.

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.

2026 is a decisive year. States have a critical opportunity, at the Seventh Review Conference of the Convention on Certain Conventional Weapons (CCW), on 16–20 November 2026, to move from discussion to negotiation on new international rules to regulate autonomous weapons systems. Existing principles of international humanitarian law apply to autonomous weapons, but new rules can clarify limits before problematic practices become entrenched.

The ICRC is calling on states to negotiate a legally binding international instrument that includes: 

The prohibition of  two types of autonomous weapons systems: 

• Those that are unpredictable in their effects, which is likely to be the case when systems incorporate forms of AI that can produce different outputs in the same circumstances.
• Those designed to target humans directly (anti-personnel autonomous weapons systems)

And restrictions on all other autonomous weapons, including limiting the types of targets to military objectives by nature, limiting the geographic scope, duration, situations and scale of their use, and incorporating effective means of neutralisation, deactivation or destruction.

The ICRC considers that these regulations are critical to addressing the humanitarian, legal, and ethical risks posed by AWS. We are not calling for a prohibition on all AWS or on AI in weapon systems altogether, as we recognize that AI might be useful in some functions such as navigation for weapons to improve accuracy.