Author: Mahi Sinha, Lloyd School of Law
Introduction:
The proliferation of autonomous and semi autonomous artificial intelligence systems has precipitated a jurisprudential quandary that traditional criminal law jurisprudence is ill-equipped to resolve. Criminal liability, since time immemorial, has rested upon the twin pillars of actus reus and mens rea- a culpable act conjoined with a blameworthy mental state. The anthropocentric framework presupposes a sentient agent capable of increasing sophistication, disrupts this paradigm by performing acts that produce harmful consequences without possessing the juridical personhood or the psychological faculties traditionally requisite for criminal culpability. This article endeavours extant models proposed by jurists and legislatures and evaluate their efficacy in a rapidly evolving technological milieu.
The conundrum of Mens Rea in Algorithmic Conduct:
The bedrock principle actus non facit reum nisi mens sit rea- an act does not render one guilty unless the mind is also guilty- becomes conceptually untenable when applied to non-sentient computational systems. An algorithm, however sophisticated its neutral architecture, lacks consciousness, intentionally, and capacity for moral reasoning, it cannot harbour malice aforethought, nor can it be said to act with recklessness or negligence in the phenomenological sense contemplated by criminal jurisprudence. The machine merely executes probabilistic computations derived from training data and optimisation functions.
The lacuna has engendered what scholars term “responsibility gap”- a situation wherein harmful outcomes are casually attributable to an AI system, yet no natural or juristic person can be conclusively identified as possessing the requisite culpable mental state. The gap is particularly pronounced in the context of deep learning systems employing opaque, non-interpretable decision making processes, colloquially termed the “ black box’” phenomenon, wherein even the system’s developers cannot fully reconstruct the causal chain leading to a particular output.
Extent models of attribution:
Jurists have proposed several doctrinal frameworks to bridge this responsibility gap. Each borrowing analogically from established principles of vicarious and corporate criminal liability.
The perpetration-via-another model treats the AI system as an innocent agent, analogous to a child or a person of unsound mind employed as an instrumentality of crime. Under this schema, liability is imputed to the programmer or user who possessed the requisite mens rea and utilised the AI as a mere conduit for effectuating their criminal design. This model retains doctrinal coherence where the harmful output was substantially foreseeable and intended by the human actor, but proves inadequate where the AI’s conduct diverges materially from its designer’s contemplation- as is increasingly common with self modifying, reinforcement-learning architectures.
The natural probable consequence doctrine extends liability to programmers or deployers where the criminal conduct, though not specifically intended, was a foreseeable by- product of the systems’s designed function. This model finds resonance with the doctrine with the doctrine of constructive liability recognised under section 3(5) of Bhartiya Nyaya Sanhita (BNS) that recognises the principle of constructive liability via common intention. However, its application to AI raises the vexed question of forseeability’s outer limits- how attenuated may the causal chain become before liability becomes unjustly remote?
The direct liability model, the most radical of the three, posits that the AI system itself ought to bear liability, necessitating a wholesale reconceptualisation of legal personhood. This model draws inspiration from corporate criminal liability jurisprudence, wherein a juristic person devoid of a physical body or singular consciousness is nonetheless subjected to punitive sanction through the doctrine of identification ( attributing the mens rea of the “ directing mind” to the corporation) or through vicarious liability principles. Proponents argue that an “ electronic personhood”, conferred with limited legal capacity, could similarly permit sanctions- fines decommissioning, or algorithmic modification- against the AI entity itself. Critics, however, contended that punishment devoid retributive or deterrent efficacy against a non-sentient entity is philosophically incoherent, as the classical penological justifications of retributions, deterrence, and rehabilitation presuppose a subject capable of experiencing punishment as such.
Comparative and statutory perspectives
The European Union’s approach, epitomised by the Artificial Intelligence Act and earlier deliberations of the European Parliament’s Committee on Legal Affairs, has gravitated towards a risk-based regulatory architecture rather than outright criminalisation of algorithmic conduct, imposing obligations of transparency, human oversight, and conformity assessment upon developers and deployers of “high-risk” AI systems. This regulatory-preventive model sidesteps the metaphysical difficulties of imputing mens rea to machines by instead constructing an ex ante compliance regime, with criminal sanctions reserved for human actors who negligently or deliberately circumvent statutory safeguards.
In the United Kingdom, the absence of bespoke legislation has necessitated reliance upon existing doctrines of gross negligence and corporate manslaughter, extending liability to entities that deploy AI systems without adequate risk mitigation. The law commission has periodically flagged the inadequacy of the Interpretation Act’s anthropocentric definitions of “person” to accommodate algorithmic agency.
Within the Indian context, the Bhartiya Nyaya Sanhita, akin to its predecessor, retains an unequivocally human-centric conception of criminal responsibility. “ person” encompasses any company, association, or body of persons, whether incorporated or not- a definition that, whilst expansive, does not readily extend to an autonomous algorithmic system devoid of incorporation or juristic recognition. Indian courts have, however, exhibited jurisprudential dexterity in analogous contexts, as evidenced by the Punjab and and Haryana High Court’s controversial 2018 pronouncement conferring “next friend” status upon an animal deity, illustrating a judicial willingness, in appropriate circumstances to expand categories of subject hood. Whether such interpretative elasticity might be extrapolated to artificial intelligence remains, as yet, jurisprudentially untested.
The problem of autonomous weapons and strict liability alternatives
A particularly exigent species of this debate arises in the context of lethal autonomous weapons systems (LAWS) and autonomous vehicles, where algorithmic decision-making directly implicates human life. Here, several commentators advocate for a strict liability regime analogous to that governing ultra-hazardous activities, wherein liability attaches to the deploying entity irrespective of fault, given the inherently unpredictable risk profile of the technology. This approach, rooted in the rule enunciated in Rylands v. Fletcher and subsequently adapted by the Indian supreme court in M.C Mehta v. Union of India through the doctrine of absolute liability, circumvents the evidentiary morass of establishing subjective fault whilst ensuring that victims of algorithmic harm are not left remediless.
Nonetheless, strict liability regimes, whilst pragmatically expedient, arguably dilute the deterrent and expressive functions of criminal law, reducing culpability determinations to a mechanical calculus of causation rather than a considered assessment of moral blameworthiness- a consequence some scholars find antithetical to the retributive foundations of penal jurisprudence.
Towards a hybrid framework
Given the inherent limitations of each isolated model, contemporary scholarship increasingly favours a hybridised, layered approach. Such as framework would retain human accountability as the primary locus of criminal sanction- targeting programmers, deployers, and institutional oversees who exercise substantive control over an AI system’s design and deployment parameters- whilst simultaneously instituting robust regulatory and civil mechanisms, including mandatory algorithmic auditing, explainability requirements, and compulsory insurance schemes, to address residual harms falling outside the ambit of individual culpability. Concomitantly, circumscribed category of “quasi-legal personhood” might be judiciously extended to sufficiently autonomous systems, not for punitive purposes per se, but to facilitate asset-based remedial mechanisms, such as compensation funds financed through mandatory algorithmic liability insurance.
Conclusion
The question of criminal liability for artificial intelligence remains, at present, an unresolved frontier straddling the disciplines of criminal jurisprudence, philosophy of mind, and technology law. Existing doctrinal frameworks- perpetration-via-another, natural probable consequence, and directly liability- each offer partial, imperfect solutions, constrained respectively for forgeability limitations, causal attenuation, and the philosophical incoherence of pushing non-sentient entities. A pragmatic resolution likely lies not in forcing artificial intelligence into pre-existing anthropocentric categories, but in forcing artificial intelligence into pre-existing anthropocentric categories, but in forcing artificial intelligence into pre existing anthropocentric categories, but in constructing a bespoke, multi-tiered accountability architecture that harmonises human culpability, regulatory oversight, and compensatory justice. As legislatures across jurisdictions grapple with this challenge, the imperative remains clear: the law must evolve with sufficient alacrity to ensure that technological innovation does not outpace the availability of legal redress for those aggrieved by algorithmic harm. The factors like mens rea and actus reus play a very critical role in finalising the guilt of a person and that is why this entire concepthas always been interpreted in a form that need someone who is being held accountable, capable of possessing the cognitive faculties to formulate these actions and the intent behind it in the first place. So whether AI can be held liable or not remains a very big question for the entire justice system. It also includes the exclamation that whether by the time the factors put under consideration should be advanced or some factors are true in essence that their credibility and essence cannot be compromised.
FAQ’s
1. Can an AI system be held criminally liable under current law?
No. Criminal liability presupposes actus reus and mens rea- a guilty act paired with a guilty mind. AI systems, however advanced, lack consciousness and moral agency, so no jurisdiction currently recognises them as criminally culpable subjects. Liability is instead attributed to human or corporate actors connected to the system.
2. What is the “ responsibility gap” in AI criminal law?
It refers to situations where an AI system causes clear, attributable harm, but no human or corporate actor can be shown to possess the requisite culpable mental state- leaving victims without a clearly identifiable wrongdoer under traditional doctrine.
