Human-like police robots are not realistic or ethically responsible,

Human-like police robots capable of replacing officers in open-ended public encounters cannot be built, according to researchers, who argue that policing requires social understanding, moral judgment, discretionary authority, and accountability capacities that current and future AI systems cannot realistically possess.

Their paper, titled Why police robots will not become human-like: social interaction, machine ethics, and limits of AI, was published in AI & Society. The study challenges growing policy and technology discussions that frame AI-driven law enforcement robots as an emerging operational reality, arguing instead that robots may assist police in narrow tasks but cannot become role-equivalent agents in socially complex policing.

Policing requires judgment that cannot be reduced to code

Advocates point to officer safety, reduced risk in violent encounters, lower fatigue, consistent enforcement, reliable recording, and improved operational efficiency. Robots do not tire, can be deployed in hazardous situations, can monitor controlled areas, and can follow de-escalation scripts without emotional escalation.

However, the authors argue that those advantages apply mainly to narrow and controlled situations. The problem begins when robots are expected to act like officers in open public life, where the meaning of an encounter is rarely fixed in advance.

Policing depends on interactional competence. Officers must interpret speech, silence, fear, hostility, confusion, deception, intoxication, trauma, and resistance in real time. A routine traffic stop can quickly become a medical emergency, a mental health crisis, a language-barrier problem, or a high-risk confrontation. The same gesture can indicate aggression, panic, compliance, confusion, or distress depending on context.

According to the authors, AI systems do not master this kind of human dialogue. Large language models can produce fluent responses, but fluency is not the same as pragmatic understanding. A system may generate convincing language without reliably understanding intentions, social norms, power relations, fear, sarcasm, or the practical consequences of an utterance in a tense encounter.

This limitation becomes critical in policing because errors can trigger coercion. A robot that misidentifies an object as a weapon, misreads a gesture as threatening, or infers hostile intent from incomplete cues could escalate a situation unjustly. Human officers also make serious mistakes, but the authors argue that robotic errors are structured differently. They may arise from sensor limits, training data, model design, software architecture, operator reliance, or post-incident interpretation, making responsibility harder to locate.

The paper also challenges the idea that more data, sensors, or larger AI models will solve the problem. In social encounters, the central difficulty is not just data capture. It is knowing which details matter and how their meaning changes as the interaction unfolds. A person’s tone, hesitation, posture, setting, history, and relationship to police may all change the meaning of the same words or actions. These factors cannot be fully anticipated or modeled in advance.

For the authors, human-like policing would require robots to decide when not to enforce a rule, when to warn rather than punish, how much force is proportionate, and how to revise actions as circumstances change. These are not merely engineering tasks. They are forms of human judgment embedded in institutions, law, training, public scrutiny, and shared social norms.

Machine ethics cannot supply police-level moral competence

Some AI ethics frameworks ask whether machines can be designed to follow moral rules, apply constraints, or make ethically informed choices. the authors argue that such approaches may work only in limited domains where the relevant situations can be clearly represented and action choices are tightly defined.

Policing is different as it involves moral judgment in unpredictable settings where values often collide. Safety, dignity, proportionality, compassion, equality, authority, and restraint may all matter at once. A police officer may need to decide whether a person is dangerous, frightened, mentally distressed, confused, defiant, or unable to comply. These judgments are rarely reducible to formal rules or training examples.

The authors also argue that legal rules alone cannot capture the practice of policing. Officers do not merely apply law mechanically. They decide what to notice, how to interpret compliance, when to intervene, when to delay action, and when a less punitive response is more appropriate. This discretionary function is central to policing, even if it is sometimes misused by humans.

A robot may apply a rule more consistently than a person, but consistency is not the same as fairness. In some cases, rigid consistency could produce systematic injustice. A robot that treats every rule violation the same way may fail to recognize vulnerability, emergency, fear, coercion, misunderstanding, or the broader social meaning of the event.

The authors state that policing legitimacy also depends on procedural justice: whether citizens believe they were heard, treated respectfully, given a meaningful opportunity to respond, and dealt with fairly. A robot may record an interaction, issue a command, or provide a scripted explanation, but that does not mean the encounter will be experienced as fair or legitimate.

Making police robots appear more human-like could worsen the problem, the paper warns. Human-like design may make a robot seem approachable, but appearance should not be mistaken for understanding. A machine that looks socially competent may encourage misplaced trust in systems that cannot actually perform human social and moral reasoning.

The study also raises concerns about AI hallucinations and model error. Even if future systems reduce such failures, the authors argue that hallucinations cannot be treated as a minor technical nuisance in law enforcement. In coercive encounters, a confident but false output can become a trigger for force. That risk is especially serious because police robots would operate in high-stakes environments where mistaken interpretation can lead to injury, unlawful detention, or death.

Accountability gaps make autonomous police robots a dangerous goal

Even if a robot could perform some policing tasks, the study warns that autonomous systems would create accountability gaps. When a human officer makes a mistake, investigators can examine training, bias, judgment, discipline, perception, decision-making, and command responsibility. When an AI-guided robot makes a mistake, blame may be dispersed across designers, data providers, software developers, police agencies, operators, procurement officials, and supervisors. This diffusion can make it harder for citizens to contest decisions or seek redress.

The paper draws attention to the risk of what accountability scholars describe as responsibility gaps and moral crumple zones. In complex automated systems, responsibility may collapse onto the nearest human operator, even if that person had limited control over the system’s decision. At the same time, institutions may use the machine’s recommendation as a shield, presenting harmful action as the result of objective technical judgment rather than human choice.

This is particularly dangerous in policing because the state’s coercive power must remain contestable and accountable. A robot that uses force, recommends force, or shapes an officer’s decision can blur the line between human authority and machine output. The authors warn that AI advice could make police violence appear more neutral, technical, or inevitable, even when the underlying action remains morally and legally contested.

The paper compares the issue to debates over automated judging, but argues that policing is even more difficult in one respect. Judicial decisions usually occur after facts are assembled through legal procedure. Policing often happens in real time, before the facts are clear, under uncertainty, emotional pressure, and possible danger.

Against this backdrop, the authors reject the goal of autonomous social policing. They argue that the appropriate path is narrow robotics under strict constraints. Police robots may be ethically justified when they reduce serious risk in bounded situations, preserve human command, maintain proportionality, allow contestation, and keep responsibility clearly assigned.

Examples include teleoperated robots for bomb disposal, robot platforms for hazardous reconnaissance, controlled logistics support, and limited sensing tools used under human supervision. These systems are not treated as replacements for police judgment. They remain tools directed by accountable human officers.