Authored By: Rishi Patel
University of Leicester
Introduction
Artificial Intelligence (AI) is no longer confined to theoretical discourse or experimental laboratories. Autonomous vehicles, predictive policing software, algorithmic financial trading systems, and medical diagnostic tools now play a central role in modern governance and commerce. As AI systems become increasingly autonomous, complex legal questions arise regarding responsibility and accountability. When an AI system causes harm—whether through a malfunctioning self-driving car or biased algorithmic sentencing tool—who should bear criminal liability?
Traditional criminal law is built upon two foundational elements: actus reus (the guilty act) and mens rea (the guilty mind). However, autonomous systems challenge this framework because AI systems can make decisions without direct human input now of harm. This article examines the difficulties AI poses to criminal liability, analyses current approaches to attribution of responsibility, and proposes a framework grounded in doctrinal coherence and policy considerations.
The Traditional Framework of Criminal Liability
Criminal liability requires proof of both a prohibited act and a culpable mental state. The House of Lords in R v Prince established that criminal responsibility ordinarily requires fault unless Parliament clearly intends otherwise.¹ The doctrine of mens rea ensures that punishment is morally justified and proportionate.
Theories of criminal liability traditionally revolve around individual agency. As HLA Hart argued, punishment presupposes that the accused had the capacity to choose otherwise.² AI systems, however, operate through machine learning processes that may evolve beyond their programmers’ specific intentions. This complicates the identification of a “guilty mind.”
The issue becomes particularly acute in cases involving autonomous vehicles. Suppose a self-driving car misinterprets sensor data and fatally injures a pedestrian. The vehicle’s decision-making algorithm may not have been directly controlled by any human at that moment. Can criminal fault be meaningfully attributed to the programmer, the manufacturer, the owner, or the corporation?
The Challenge of Autonomous Decision-Making
Modern AI systems employ deep learning models that adapt through exposure to data. Unlike traditional software, which follows explicit instructions, machine learning systems generate outcomes through probabilistic analysis. The opacity of such systems—often referred to as the “black box” problem—makes foreseeability difficult to assess.
Foreseeability is central to negligence and recklessness. In R v G and Another, the House of Lords redefined recklessness as requiring subjective awareness of risk.³ If no individual specifically foresaw the precise harmful output of an AI system, establishing subjective recklessness becomes problematic.
Furthermore, the autonomy of AI systems raises philosophical concerns. While AI does not possess consciousness or moral agency, it can perform tasks independently. However, criminal law does not recognise non-human entities as moral agents capable of forming mens rea (with limited exceptions such as corporate liability). Therefore, liability must ultimately revert to human actors or legal persons.
Corporate Criminal Liability as a Model
One possible solution lies in adapting principles of corporate criminal liability. Corporations, like AI systems, lack human consciousness. Yet the law attributes criminal responsibility to them through doctrines such as identification theory and vicarious liability.
Under the identification doctrine, as articulated in Tesco Supermarkets Ltd v Nattrass, liability is attributed to a corporation when the “directing mind and will” of the company commits the offence.⁴ This model may be insufficient for AI-related harms because harm may arise from systemic programming decisions rather than a single controlling individual.
The Corporate Manslaughter and Corporate Homicide Act 2007 offers a more flexible approach by focusing on management failures rather than individual culpability.⁵ A similar framework could be extended to AI developers and companies deploying autonomous systems. Liability could arise where systemic design flaws, inadequate testing, or failure to implement safety protocols amount to a gross breach of a duty of care.
Programmer and Manufacturer Liability
Another approach involves assigning liability to programmers or manufacturers under negligence principles. If harm results from foreseeable risks that were inadequately addressed, criminal negligence may be established.
However, over-criminalising developers risks stifling innovation. AI systems often function in unpredictable environments, and not every harmful outcome indicates fault. Criminal law must balance deterrence with fairness. As Andrew Ashworth argues, criminal liability should be reserved for conduct demonstrating sufficient culpability and moral blameworthiness.⁶
Imposing strict liability for AI-related harm would undermine the foundational principle that criminal punishment requires fault. Strict liability may be appropriate in regulatory contexts, but its expansion into serious offences would be controversial.
The Case for a “Responsible AI Operator” Model
A more balanced solution may lie in adopting a “Responsible AI Operator” model. Under this framework:
- The entity deploying the AI system (corporation or individual) would bear primary responsibility.
- Liability would depend on failure to implement reasonable safeguards.
- A due diligence defence would be available where robust risk mitigation measures were in place.
This approach mirrors regulatory models in areas such as data protection. Under the UK General Data Protection Regulation (UK GDPR), organisations must implement appropriate technical and organisational measures to ensure compliance.⁷ A similar compliance-based framework for AI could incentivise proactive risk management.
Such a model aligns with emerging international developments. The European Union’s proposed AI regulatory framework adopts a risk-based approach, categorising AI systems according to potential harm and imposing corresponding obligations.⁸ Integrating criminal liability into this structure would ensure coherence between regulatory and penal responses.
The Limits of Granting AI Legal Personality
Some scholars have suggested granting AI systems limited legal personality. However, this proposal remains controversial. Legal personality without moral agency creates conceptual tension. Unlike corporations, which are associations of natural persons, AI systems are technological artefacts.
Granting AI criminal liability would not meaningfully advance deterrence or retribution, as AI systems cannot experience punishment. The communicative function of criminal law—expressing societal condemnation—requires a human or collective subject capable of moral understanding.
Therefore, responsibility must ultimately rest with human agents or corporate entities.
Policy Considerations
Any reform must consider the following policy factors:
- Fair Labelling
Criminal law should accurately reflect the nature of wrongdoing. Labelling a developer as a “manslaughter offender” for unforeseeable AI malfunction risks disproportionate stigma.
- Deterrence
Liability should incentivise robust testing, transparency, and monitoring of AI systems.
- Innovation
Overly punitive regimes may discourage beneficial technological advancement.
- Public Confidence
Clear accountability structures are essential to maintain trust in emerging technologies.
Balancing these objectives requires a nuanced approach that neither absolves responsibility nor imposes unjust blame.
Proposed Framework for Reform
A coherent criminal law framework for AI should incorporate:
- Statutory clarification defining the duties of AI developers and operators.
- Gross negligence threshold for serious criminal offences.
- Corporate-focused liability rather than individual programmers in most cases.
- Mandatory compliance standards tied to safety audits and transparency.
- Due diligence defences where reasonable precautions were taken.
Such reforms would preserve core criminal law principles while adapting to technological realities.
Conclusion
Artificial intelligence presents one of the most significant challenges to modern criminal law. The traditional requirement of mens rea presupposes human cognition and moral agency, yet autonomous systems blur the link between action and intent. Granting AI systems an independent criminal personality is neither doctrinally coherent nor practically effective.
Instead, liability should focus on the human and corporate actors responsible for designing, deploying, and supervising AI systems. A compliance-based, risk-oriented framework—supported by a gross negligence threshold—offers the most balanced solution. By preserving fundamental principles of fault while adapting to technological innovation, criminal law can remain both just and relevant in the age of artificial intelligence.
Footnote(S) (OSCOLA Style)
- R v Prince (1875) LR 2 CCR 154.
- HLA Hart, Punishment and Responsibility (2nd edn, OUP 2008).
- R v G and Another [2003] UKHL 50, [2004] 1 AC 1034.
- Tesco Supermarkets Ltd v Nattrass [1972] AC 153 (HL).
- Corporate Manslaughter and Corporate Homicide Act 2007.
- Andrew Ashworth, Principles of Criminal Law (9th edn, OUP 2019).
- UK General Data Protection Regulation, art 24.
- Proposal for a Regulation of the European Parliament and of the Council Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) COM/2021/206 final.
