Updated 2026-09-12
from the news desk
Can AI be charged with a crime? Why criminal law still looks for a human
Jaroslav Fenyk’s 2026 paper rejects direct criminal liability for AI systems, while John Kingston’s 2018 paper discusses a narrower possibility in strict-liability offences. In both accounts, liability analysis still centres on the people…
On the legal scholarship in these sources, AI is not a straightforward target for criminal charges. Jaroslav Fenyk rejects direct criminal liability for AI systems, while John Kingston describes a narrower argument that some strict-liability offences could be attached to AI conduct; in both accounts, the main legal focus remains on the humans behind the system.
Why the short answer is no
Fenyk’s paper, which examines criminal liability through a Czech criminal law perspective, states the core conclusion directly: “The conclusions of the paper are negative with regard to the possible direct liability of such systems.”
Fenyk’s abstract also gives his proposed answer to the liability gap. He writes: “A solution is proposed in the form of criminal liability for persons who create, control and monitor autonomous and non-autonomous systems.”
That makes the plain-English answer fairly short. On Fenyk’s view, if an AI-linked event reaches criminal law, the system itself is not the proper bearer of direct criminal liability. The people around it remain the relevant legal actors.
Kingston’s paper is more exploratory. Kingston says his paper asks whether criminal liability could ever apply to AI systems, and to whom. Even so, Kingston’s discussion still spends much of its time tracing liability back to programmers, users, vendors or service providers. So the two papers do not say exactly the same thing, but they do point in the same practical direction: criminal law still searches for a human role.
What actus reus, mens rea and will mean here
Kingston says criminal laws normally require both an actus reus and a mens rea. In plain English, Kingston treats actus reus as the prohibited act or failure to act, and mens rea as the required mental state, such as knowledge or negligence.
Kingston says Hallevy proposes three models for AI-related offences, and Kingston summarises that framework as follows: “Hallevy goes on to propose three legal models by which offences committed by AI systems might be considered.”
Fenyk frames the issue from another angle. His abstract says the paper analyses the place of human will in the foundations of criminal liability and compares that requirement with autonomous and non-autonomous systems equipped with AI. On that framing, the obstacle is not only whether a machine caused an act. It is also whether criminal liability, as traditionally understood, presupposes a kind of human willing that AI does not have.
Kingston’s paper makes a similar point in the language of mens rea. He writes: “Assigning a mens rea is much harder, and so it is here that the three levels of mens rea become important.”
That matters because it shows why this is not only a technology question. It is also a legal-person question. A separate explainer on this site covers legal personhood, and another explains electronic personhood. The criminal-law issue here is narrower: even if a system acts in the world, can the law treat it as having the kind of imputable mental state that criminal punishment normally requires?
Why these accounts still place liability on humans
Kingston’s first model, drawn from Hallevy, treats the AI as an innocent agent. Kingston says that where an offence is carried out by an entity that lacks the mental capacity to form mens rea, liability can attach to the person who instructed or used that entity. Applied to AI, Kingston says the programmer or user could then be treated as the perpetrator-via-another.
Kingston’s second model is the natural-probable-consequence approach. In Kingston’s account, this is the case where part of a program built for a legitimate purpose is activated inappropriately and produces a criminal act. Kingston says users or, more probably, programmers might then be held liable if they knew that a criminal offence was a natural and probable consequence of the program or its use.
Fenyk’s conclusion fits that human-centred logic. His proposed solution is liability for the persons who create, control and monitor the system. That does not answer every allocation question between designer, programmer, deployer and operator. But it does keep the unit of criminal responsibility human.
One reading of both papers is that criminal law is doing two jobs at once. It is trying to explain harmful machine conduct, and it is trying to preserve doctrines built around human agency. That helps explain why a separate site explainer on whether AI can sign a contract also ends up asking who the legally recognised person is behind the automated act.
The narrower argument for direct AI liability
Kingston does not rule out direct AI liability across the board. He says it is relatively simple to attribute actus reus to an AI system when the system takes an action that results in a criminal act, or fails to act when there is a duty to act.
The harder issue, Kingston says, is mens rea. That is why his strongest opening for direct liability is not the whole criminal code, but offences that demand less proof of mental state. Kingston says strict-liability offences do not require intent to be proved, and he gives a concrete illustration from self-driving cars. As Kingston puts it: “If a self-driving car was found to be breaking the speed limit for the road it is on, the law may well assign criminal liability to the AI program that was driving the car at that time.”
Kingston’s conclusion pushes the point a bit further. He says it seems unlikely that AI programs would contravene laws requiring knowledge that a criminal act was being committed, but very possible that they might contravene laws using a reasonable-person form of negligence, and almost certain that they could contravene strict-liability offences.
The steelman for direct liability therefore exists, but it is narrow. It rests less on the claim that AI has a human-like guilty mind, and more on the claim that some offences ask the law to prove much less of a mind in the first place.
What would have to change for AI to be charged directly
These papers suggest that direct AI charging would need more than proof that software caused harm. Fenyk’s paper indicates that, at least on his account of criminal liability, the missing piece is will. Kingston’s paper indicates that the missing pieces also include workable rules on mens rea, defences and punishment.
Kingston notes that once an AI system is treated as directly liable, further questions follow. He asks whether a malfunctioning program could claim a defence analogous to insanity, whether malware interference could resemble coercion or intoxication, and who or what would actually be punished. Kingston presents those as open problems, not settled doctrine.
So the strongest answer available from these sources is limited but stable. If a reader asks whether AI can be charged with a crime in the same straightforward sense as a human offender, Fenyk’s answer is no. If a reader asks whether legal theory has explored narrower exceptions, Kingston’s answer is yes, especially for strict-liability offences. In both lines of analysis, however, criminal law still does most of its work by locating a responsible human being.
Frequently asked questions
Does any source here say AI can never be criminally liable?
Not in universal terms. Fenyk rejects direct criminal liability for AI systems, but Kingston discusses models under which AI might be treated as directly liable, especially for strict-liability offences. The disagreement is about how far that possibility goes.
What is mens rea, and why does it matter for AI?
Kingston says mens rea is the required mental element of an offence, such as knowledge or negligence. It matters because Kingston says assigning mens rea to AI is much harder than assigning the act element, which is one reason direct AI liability is limited in his account.
If the AI is not charged, who might be?
Fenyk proposes liability for the people who create, control and monitor AI systems. Kingston says liability might also fall on the programmer, the user, the vendor or the service provider, depending on which legal model applies and what could be foreseen.
Are strict-liability offences different?
Yes, in Kingston’s analysis. Kingston says strict-liability offences do not require proof of mens rea, and that makes them the clearest opening in his paper for direct AI criminal liability.
caveats: The broad no-direct-liability answer relies mainly on Fenyk’s 2026 abstract. The three-model framework, the self-driving speeding example, and the discussion of mens rea, defences, punishment and human-side allocation questions come from Kingston’s 2018 paper alone.