ABSTRACT

1 INTRODUCTION

Generative Artificial Intelligence, led today by models such as ChatGPT, represents a pivotal milestone in the trajectory of global technological transformation. It is no longer merely an executive tool but has become a cognitive agent actively participating in content creation, problem-solving, and decision-making. This shift is clear in the growing reliance on AI systems across multiple fields—including education, healthcare and law—posing unprecedented challenges to legal systems, particularly the criminal justice system. The system now faces uncertainty in identifying the “actor” in cases involving harm or criminalised conduct resulting from AI intervention.

From this standpoint, the present study, titled ChatGPT in the Dock: Reflections on the Future of Criminal Liability, aims to analyse the legal dilemmas posed by this model, one of the most prominent and controversial manifestations of generative artificial intelligence. Its primary objectives are: (1) to expose the shortcomings of traditional concepts of criminal liability in light of AI intervention; (2) to analyse the complex structure of liability resulting from the actions and outputs of systems like ChatGPT; and (3) to propose restructuring legal attribution rules in a manner that ensures the protection of individual rights and the effectiveness of criminal deterrence in the age of intelligent machines.

Upon AI autonomy, the study takes a two-pronged approach: theoretically, it examines criminal concepts such as intent, causation, and actor liability; practically, it examines comparative legal models, particularly the 2024 European AI Act, and their applicability to cases involving harmful or deceptive AI use, like ChatGPT. The study is grounded in a central hypothesis: current frameworks of criminal accountability are incapable of comprehending the outputs of partially autonomous non-human entities. This necessitates the development of more flexible legal models, such as collective or virtual liability, or even the establishment of a new concept: the "artificial actor."

The core legal questions this study addresses include: Who bears criminal liability when ChatGPT use results in harmful or criminal outputs? Should liability be limited to the end user, or should it extend to developers, operators, and designers? To what extent can ChatGPT be considered an independent actor contributing to the criminal outcome? What is the optimal legislative pathway to bridge the legal gap caused by the ambiguity of intent and discernment in AI-generated actions? These questions are not merely theoretical; they lie at the heart of the challenges facing criminal justice in the coming decades. Exploring them constitutes a legal and strategic necessity for ensuring the sustainability of the judicial system in a rapidly evolving digital environment.

This study on criminal liability for AI systems like ChatGPT addresses the shortcomings of existing legal frameworks, identifies the parties liable, and the challenges in assessing intent. It examines legitimate risks, regulatory gaps, and the difficulties of applying traditional criminal law to autonomous AI behaviour. Finally, it proposes future-oriented solutions to change liability frameworks, such as corporate responsibility and systemic accountability.

2 LITERATURE REVIEW

The scientific literature indicates that criminal accountability for artificial intelligence systems, particularly complex software such as ChatGPT, faces fundamental challenges related to the nature of algorithmic actions and the opacity of decision-making processes.¹ Many researchers point out that AI operates through intricate algorithmic networks beyond human control, making it more difficult to attribute any resulting harm to conventional ideas of fault or intent.² Traditional criminal liability, which is based on direct human action, is being reexamined amid the crisis of causal attribution.³

In the same context, recent legal studies have addressed the crisis of lack of control over the technical risks posed by AI, illustrating that traditional laws are incapable of regulating these new technological risks within conventional liability frameworks.⁴ To

address risks posed by software like ChatGPT, the legal landscape is moving toward preventive and abstract liability models, emphasising producers' accountability for technical safety measures rather than for direct harm.⁵

Moreover, the literature highlights legislative and practical challenges stemming from the exclusion of intelligent software from traditional product laws, particularly those that define a product solely as a physical entity, thereby obstructing effective accountability for such systems.⁶ In this regard, the European Artificial Intelligence Act emerges as an advanced model imposing deterrent sanctions on violations and promoting enhanced transparency and disclosure of AI-related risks.⁷ Recent studies also recommend integrating administrative oversight with criminal liability, thereby opening new horizons for managing technological risks through an integrated and evolving legal framework that safeguards fundamental rights without impeding technological innovation.⁸

3 METHODOLOGY

This study aims to analyse the legal framework governing criminal liability for generative artificial intelligence systems, focusing on the ChatGPT model as a contemporary practical example that reflects the legal and technical challenges in this field. To achieve this, the study employs a descriptive-analytical approach that examines national and international legislative texts and relevant legal principles, and reviews contemporary jurisprudential and legal literature on the nature of criminal liability amid the rapid development of AI technologies. Additionally, the study employs a comparative legal method, which systematically identifies and investigates specific areas where legal systems diverge and converge. The comparative criteria include factors such as the degree of criminal liability, regulatory protection, enforcement tactics, and ethical accountability standards. Using this framework, the study contrasts advanced European legislation, particularly the Artificial

Intelligence Act (AI Act),⁹ with local legislative systems that are still in their infancy in addressing emerging technological challenges.¹⁰

The study, employing a critical approach, assesses the effectiveness of current legal frameworks in addressing AI-related harms, emphasising the challenges of establishing traditional causal links between actions and outcomes in algorithmic contexts. It explores the potential adoption of new liability models centred on preventive liability and criminal negligence, using case studies to illustrate how laws can evolve to balance societal protection with the promotion of technological innovation.

4 CRIMINAL LIABILITY DEFINITION FOR ARTIFICIAL MINDS AND CHATGPT

4.1. A Mind Without a Body: Who Prosecutes ChatGPT?

The phenomenon of ChatGPT vividly exemplifies the profound complexities artificial intelligence introduces into the criminal legal system.¹¹ This advanced linguistic system does not merely process data, but generates textual decisions that interact with humans and influence their cognitive, social, and legal realities.¹² When the generated text becomes capable of shaping convictions or guiding decisions, we are no longer dealing with a mere silent technical tool but a virtual mind without a body—one that redefines legal agency.¹³ This raises fundamental questions about the nature and legal classification of artificial intelligence, especially in the absence of a unified definition within legal systems.¹⁴ Traditional criminal models of intent, perpetrator identification, and liability must be reevaluated as AI's legal identity straddles the line between a human-controlled tool and an autonomous decision-maker.¹⁵

When we ask, "Who prosecutes ChatGPT?", we pierce through the veil of classical law and enter an unprecedented legal space.¹⁶ The European Commission's 2021 proposal,¹⁷ despite its effort to define AI broadly as systems capable of "generating outputs that affect the environment," fails to address the core dilemma: how to distinguish between basic AI and those complex generative systems that produce socially and legally impactful texts—like ChatGPT.¹⁸ The challenge lies in three central characteristics: autonomy, interactivity, and opacity.¹⁹ The term "autonomy" describes AI's capacity to act without direct oversight, not its intention. While opacity reflects the "black box" nature of its unpredictable algorithms, interactivity arises from daily user engagement.²⁰ These traits complicate the attribution of criminal liability, as the lines blur between programming error and human intention, between spontaneous output and directed decision.²¹

The dilemma posed by ChatGPT goes beyond legal debate into deep philosophical and ethical territory, shaking the foundations of traditional criminal concepts.²² The absence of premeditation, the unpredictability of outputs, and the difficulty in identifying a clear actor, be it the developer, the user, or the owning company, reshapes the question of criminal liability.²³ It transforms it from a simple binary framework into a multi-layer network.²⁴ The traditional legal system, built on the formula "actor–victim–harm," is no longer adequate to encompass intelligent entities that commit crimes not in conventional ways but through knowledge flows open to interpretation.²⁵

Therefore, there is a pressing need for a new legal model based on "distributed shared accountability" and transparent oversight mechanisms that allow for tracing algorithmic decisions and assessing their legality and impact.²⁶ Ultimately, prosecuting a “mind without a body” is not a metaphor—it is a tangible necessity that demands an epistemological revolution in our understanding of law and a careful balance between technological innovation and legal protection of rights and freedoms in the digital age.²⁷

4.2. When ChatGPT Speaks Without Criminal Intent: Can It Be Prosecuted?

We are not here to explore the intricate technical workings that give ChatGPT its structure, as these have been thoroughly explained by specialists and go beyond the concerns of criminal jurisprudence.²⁸ Even in cases where there is no criminal intent or will, what matters is how AI-generated linguistic outputs affect the legal system.²⁹ The machine learning model does not operate on explicit logical rules; rather, it follows a probabilistic inductive approach, extracting linguistic patterns from billions of examples without “understanding” them. ChatGPT neither knows the truth nor intends to lie, yet it can generate harmful, misleading, or inflammatory content.³⁰ AI thus generates "speech without reason" and "action without intent," undermining traditional legal theories that attribute criminal liability exclusively to human consciousness and rational awareness.³¹

When ChatGPT generates illegal or criminally consequential content, the challenge of identifying the responsible actor emerges. Is it the model itself? The developers? The owning company? Or the users? Like a "blind painter," the model uses probabilistic estimates that change as its inputs change to create linguistic portraits without knowing why or for what purpose.³² To maximise flexibility and generative capacity rather than the logical consistency required for legal liability, its architecture is deliberately opaque.³³ Under the new legal concept of distributed liability, traditional actors cannot be held

accountable; instead, accountability must be traced across multiple stakeholders, from user interfaces to training environments.³⁴

Even though ChatGPT does not aim to break the law or incite, its reliance on linguistic patterns can yield results with significant ethical or legal implications. The legal challenge lies here: how do we assign accountability to a system that lacks will, yet produces effects? The answer requires moving beyond traditional liability frameworks by developing new rules that hold developers and operators accountable and enforce proactive oversight of inputs and algorithms. The legal focus must shift from “the actor’s intent” to “design and operational responsibility,” and from the “criminal mind” to a system of “digital governance.” This model does not err because it chooses to, but because it lacks the capacity to distinguish right from wrong—necessitating legislation that redefines the relationship between technology and accountability under a new logic.³⁵

ChatGPT stands at the threshold of artificial consciousness, in a legal grey zone that criminal justice systems are not yet prepared to handle. It does not think or comprehend, but it generates discourse that simulates thought. It is neither a traditional actor nor a mere tool—it is a linguistic entity that challenges settled legal classifications. In this sense, artificial intelligence functions more as a mirror exposing the inadequacies of our laws than as a standalone problem. The challenge lies not in how “intelligent” it is, but in how legally “prepared” we are to incorporate it into our network of criminal concepts. A reevaluation of crime and punishment in which actors may be nonhuman, and liability arises from error, probability, or unanticipated consequences rather than conscious intent, would result from failing to act, risking the prosecution of algorithms for unintended outputs.

4.3. Attributing Criminal Liability in the Age of Intelligent Machines

Artificial intelligence systems pose a genuine challenge to the traditional criminal liability framework, which is built upon the pillars of actus reus (the act), mens rea (intent), and will.³⁶ These intelligent entities are neither human nor self-aware nor criminally intent; rather, they are digital tools that generate unpredictable behaviours that are difficult to foresee accurately.³⁷ As reliance on these complex systems—operating on probabilistic rather than explicit logical bases—increases, a central legal question arises: How can liability for harm caused by these systems be assigned when humans lack full control over their behaviour? ChatGPT and similar language models do not rely on true understanding. Still, on intricate statistical patterns they neither comprehend nor can

explain, making prosecution for harmful actions impossible under traditional criminal law concepts.³⁸ Hence, our legal system faces a profound crisis in how to classify these non-living entities that produce real-world effects and in defining a legal framework balancing technological progress with justice.³⁹

Confronting this dilemma, international approaches, especially within the European Union, have emphasised the importance of strict civil liability as a practical mechanism to protect victims.⁴⁰ Directive 85/374/EEC on defective products is a law that holds manufacturers liable for damages without needing proof of fault or intent.⁴¹ Traditional concepts of "defect" and "causality" are blurred by the complexity of AI since intelligent systems are dynamic entities with probabilistic behaviours rather than traditional products.⁴² This prompted the European Commission in 2022 to propose comprehensive legal updates addressing “smart products”.⁴³ These updates aim to broaden legal protection and shift the burden of proof onto producers by presuming a link between defect and damage automatically, and imposing economic responsibility on producers for risks associated with these systems— even when their behaviour is unpredictable or unforeseeable.⁴⁴ This reflects a fundamental shift in liability philosophy—from focusing on the actor’s intent to ensuring effective compensation for victims regardless of the actor’s awareness.⁴⁵

Since AI is an unconscious entity lacking intent or will, the law must transcend traditional concepts grounded in these elements and develop a hybrid legal framework combining strict civil responsibility, regulatory liability, and supervisory oversight.⁴⁶ Producers or developers oversee the implementation of safety precautions and ensure clear standards, aiming to prevent harm and provide victims with straightforward compensation.⁴⁷ Instead of prosecuting algorithms without understanding how they work, the AI approach

emphasises accountability for system design and operation, prevention, and compensation.⁴⁸ Without renewing our legal frameworks, we risk a future in which algorithms are prosecuted for unintended or unforeseen actions, opening a dangerous legal vacuum that threatens the application of justice.⁴⁹ 5 CHALLENGES IN ASSESSING THE CRIMINAL LIABILITY OF ARTIFICIAL INTELLIGENCE

5 CHALLENGES IN ASSESSING THE CRIMINAL LIABILITY OF ARTIFICIAL INTELLIGENCE

Conventional criminal law faces unprecedented challenges because of the rise of ChatGPT and other generative AI systems. At the core of this disruption is the causality dilemma: it is impossible to use conventional legal methods to link AI actions to outcomes. The fact that AI operates without conscious intent, which contradicts conventional notions of guilt, exacerbates the mens rea dilemma. When taken as a whole, these crises highlight the pressing need to reconsider liability assessment in the age of intelligent machines.

5.1. The Crisis of Criminal Law in the Era of ChatGPT

The digital revolution is causing a major crisis for criminal law, which is still confined to traditional frameworks based on human consciousness, intent, and choice.⁵⁰ The criminal system is governed by well-established principles such as legality, personal blame, and the presumption of innocence, designed to address human actions with clear motives. Still, it is ill-equipped to accommodate acts generated by generative AI systems like ChatGPT.⁵¹ Determining criminal liability in systems with self-learning capabilities and programmers' autonomy is challenging because of unpredictable behaviours that cannot be traced to specific human actors.⁵² This clash represents the "shock of modernity" described by the Italian thinker Federico Stella, in which traditional criminal concepts such as intent and awareness lose their relevance when confronted with technology characterised by behavioural ambiguity and unpredictability.⁵³

AI-induced harm challenges the evidentiary and proof mechanisms of criminal law because algorithmic biases or training errors make it more difficult to assign blame and to establish a

direct causal relationship between the action and the harm.⁵⁴ In a participatory, multi-agent technological environment, personal liability becomes unclear as multiple parties' acts and omissions intertwine, weakening courts' certainty and threatening the principle of predictability of outcomes.⁵⁵ For instance, output generated by ChatGPT may be used to incite, defame, or commit fraud, but the absence of direct human intent complicates criminal accountability and exposes the limitations of current legal tools to address these complexities.⁵⁶

The natural outcome of this crisis is a clear and troubling legal vacuum, where current criminal law lacks explicit and comprehensive rules addressing acts resulting from AI that lack awareness or will.⁵⁷ These vacuum places justice in a predicament: AI itself cannot bear criminal liability, nor can the traditional responsible human—whether developer or user— be easily held liable due to algorithmic complexity and the difficulty of proving fault and intent.⁵⁸ Consequently, today’s digital reality demands a rethinking of the very definitions of crime and the principles of liability, opening the door to profound legal and philosophical debates about the limits of criminal law and how to develop a regulatory framework that balances societal protection from AI risks with encouraging innovation without threatening legal security and justice.⁵⁹

5.2. The Causality Dilemma in the Era of Generative Artificial Intelligence

Generative AI systems like ChatGPT challenge traditional legal notions of causality.⁶⁰ Criminal law’s deterministic model linking human acts to outcomes struggles against AI’s probabilistic algorithms,⁶¹ which produce unexpected outcomes without a human actor,⁶² undermining classical proof and responsibility frameworks.⁶³

It is challenging to prove a causal link between the actions of generative AI systems and the outcomes they generate due to their technical complexity. This puts conventional legal reasoning to the test.⁶⁴ In the past, courts used instruments like digital "black boxes" that capture event data to piece together the sequence of an incident and identify its cause.⁶⁵ The algorithm's internal decision-making processes, however, are based on complex, nonlinear probabilistic models, and these tools only provide preliminary indications.⁶⁶ Experts and judges are unable to conclusively determine whether an error with ChatGPT results from a programming error, bias in the training data, or user behaviour.⁶⁷ Such ambiguities make it more difficult to establish clear causation, particularly in cases of harmful content generation, and complicate the application of conventional principles that rely on a clear connection between action and result.⁶⁸

Criminal law faces an epistemic crisis due to this causality conundrum, compelling a reexamination of its central concepts of intent, causation, and liability.⁶⁹ The nomologicaldeductive model that judicial systems employ excludes other possible explanations and necessitates precise scientific law and a logical causal relationship between an action and its result.⁷⁰ Generative AI undermines this model, as even system developers cannot precisely identify the causes of the outcomes they produce, rendering actions into multidimensional probabilistic outcomes.⁷¹ Legal systems must accept “flexible causality”, place blame on programmers, users, and AI systems, and update evidentiary standards to account for social and technical context to ensure criminal justice.⁷² Law must adapt to this new reality, avoiding confinement within rigid models that fail to accommodate the algorithmic revolution, thus preserving the essence of justice and individual rights in the age of artificial intelligence.⁷³

5.3. ChatGPT and the Mens Rea Dilemma

Amid the rapid development of generative artificial intelligence, the ChatGPT model has emerged as an icon of digital transformation in natural language processing.⁷⁴ However, this model embodies a fundamental legal crisis that touches the core of causal proof—a foundational element in determining criminal liability. Traditional legal concepts rely on a clear and direct causal relationship between a human act and a harmful result, based on a nomological model that infers harm from a specific act carried out with awareness and intent.⁷⁵ By contrast, ChatGPT, as a complex algorithmic system operating through millions of probabilities, does not produce actions with intent or purpose. Its outputs are based on data and deep learning, not on conscious decisions.⁷⁶

For example, if ChatGPT generates false or inciting content that causes moral or material damage, a complex question arises: can it be proven that the model's output directly caused these outcomes? Or is the human user who deployed the content responsible? In such cases, traditional legal rules requiring the attribution of harm to a conscious being become difficult to apply when faced with the opacity of an "artificial mind" that possesses neither will nor intention.⁷⁷

The mental element (mens rea) is a cornerstone of criminal liability, typically manifested in intent or negligence.⁷⁸ However, when discussing ChatGPT, these concepts begin to dissolve, as the model has no will or consciousness and thus cannot possess intent like a human actor.⁷⁹ This raises the question of whether producers or developers are exonerated. Not necessarily, as existing doctrines provide alternative bases for attribution, the law might invoke the notion of probable intent, holding developers or users accountable if they could reasonably foresee the harm as a potential consequence of using the model—even in the absence of malicious intent.⁸⁰

Suppose a developer or company releases a version of ChatGPT capable of generating inciting or offensive content, knowing that it could be misused, yet fails to implement sufficient preventive measures. Here, probable intent is realised, as they are expected to foresee the risk and prevent it.⁸¹ Conversely, if harm results from unexpected use within a

wide range of possible outcomes, assigning liability becomes more complex.⁸² The model’s opacity adds to the challenge, as even the developers might not be able to explain why certain content was generated, making it difficult to determine whether the act constituted criminal negligence or an acceptable technical risk.⁸³

To address these complexities, the law must go beyond mere adjustments to evidentiary tools. It must construct a new regulatory system that considers the nature of generative artificial intelligence.⁸⁴ For instance, clear legal frameworks should be established to govern the development and operation of models like ChatGPT, imposing strict standards for digital safety and governance responsibility, and defining the limits of acceptable risks—or what might be termed the “permissible risk zone”.⁸⁵

Imagine a company developing ChatGPT that implements preventive measures such as content filtering and usage management to prevent harmful outputs. These measures would delineate the permissible risk zone, serving as the dividing line between acceptable technical error and liability overreach.⁸⁶ Liability would arise only when these boundaries are breached, due to a lack of precautionary measures or willful disregard for potential risks.⁸⁷

Thus, the notions of intent and negligence must be reformulated within a broader framework that accounts for varying degrees of foreseeability and probability, while recognising the unique nature of “actions” generated by a non-conscious mathematical system.⁸⁸ The law now faces a fundamental philosophical challenge—not merely redrawing evidentiary tools but reconstructing its foundational perceptions of criminal liability in an era where artificial intelligence has become an independent, complex actor generating actions and consequences that transcend traditional concepts of will and intent.⁸⁹ 6 THE LEGAL AND REGULATORY CHALLENGES OF GENERATIVE AI

6 THE LEGAL AND REGULATORY CHALLENGES OF GENERATIVE AI

As generative artificial intelligence, like ChatGPT, grows quickly, new legal issues regarding legal risks and responsibility for generated content surface. Determining criminal liability is challenging due to a regulatory gap left by current frameworks that have not kept up with the rate of innovation. To suggest practical ways to ensure the safe and lawful use of this

technology, this theme investigates the nature of algorithmic intelligence, evaluates legal risks, and looks at regulatory gaps.

6.1. The Algorithmic Regulatory Gap and Legal Accountability of ChatGPT

Generative AI models, led by ChatGPT, are characterised by a unique cognitive nature, as they operate on complex algorithms that form what is known as the “black box,” whose inner workings are difficult to explore or whose outputs are hard to interpret precisely.⁹⁰ In the absence of a comprehensive scientific framework clarifying the causal pathways these models follow in making decisions, users face a fundamental challenge: the inability to understand how the model selects certain phrases or information over others, or to interpret the logic underlying the arrangement of elements in the response.⁹¹ For example, when ChatGPT provides a legal recommendation in a criminal case, neither the judge nor the lawyer has access to the detailed reasons that led the model to formulate this recommendation, which obstructs the possibility of accountability if an error or harm occurs.⁹² In this sense, this interpretive gap undermines trust in AI outputs and poses fundamental challenges in terms of transparency and legal accountability.⁹³

A significant portion of errors produced by models like ChatGPT stems from intertwined technical and legal factors.⁹⁴ On the technical level, some errors relate to the design of algorithmic architecture, where certain arrangements in the neural network layers can lead to unintended implicit biases, such as the model discriminating between individuals based on gender or race.⁹⁵ At the data level, AI models rely on massive amounts of text drawn from diverse online sources, which may carry cultural or ethical biases or contain inaccurate legal information, thereby transferring these biases into the model’s outputs.⁹⁶

Given these multiple challenges, there is an urgent need to develop a strict legal regulatory framework governing the operation of generative AI models and ensures the safety and reliability of their outputs, especially in highly sensitive contexts such as the legal field.⁹⁷ The European AI Act represents a pioneering step in this direction,⁹⁸ imposing on developers of models like OpenAI a comprehensive risk management obligation, requiring them to ensure training data quality, prepare technical documentation that clarifies decision-

making logic, maintain accurate logs of system outputs, and mandating human oversight of system outputs in high-risk cases.⁹⁹ Legal requirements for risk reduction or elimination could make the judiciary less accountable, calling for new models of legal accountability that account for the cognitive differences between humans and machines.¹⁰⁰ This necessitates establishing mechanisms for interpretation and analysis that assist users in consciously and thoughtfully understanding and evaluating the model’s results.

6.2. Algorithmic Intelligence and Legitimate Risks

The development of generative artificial intelligence systems, such as the ChatGPT platform, raises novel legal challenges concerning liability for damages resulting from their use.¹⁰¹ Even if the manufacturing company complies fully with all prescribed technical and regulatory standards, the pressing question remains: Is such compliance sufficient to absolve it of legal liability? Legal and philosophical experience indicates that formal adherence to rules does not necessarily prevent harm, especially when algorithms are granted quasiautonomous power to make complex decisions, as is the case with these systems operating within unpredictable environments and multifaceted causes.¹⁰² For example, an algorithmic error in interpreting a query or generating inaccurate content may cause serious psychological or social harm that cannot be fully anticipated, thereby fueling debate over whether technical standards alone provide an adequate legal defence.¹⁰³

In legal theory, compliance with rules is recognised as a necessary but insufficient condition for relieving liability, as courts also rely on the “reasonable actor” or “person of similar position and competence” standard.¹⁰⁴ This requires assessing whether the manufacturer took the necessary professional and prudent measures to avoid harm. In complex technical industries, jurisprudence acknowledges the existence of “residual risks” or “acceptable risks”—those that cannot be eliminated except at exorbitant costs or at the expense of technological progress.¹⁰⁵ According to Amy Stein,¹⁰⁶ these risks represent an implicitly accepted zone of risky behaviours, provided they remain within the bounds of control and reasonable precautions, reflecting the reality of technological development and balancing innovation with safety.

On the legislative front, the European Artificial Intelligence Act (AI Act)¹⁰⁷ exemplifies legal evolution that abandons the ideal of eliminating risks and instead adopts a pragmatic philosophy aimed at minimising risks as much as possible while acknowledging their persistent possibility.¹⁰⁸ This law not only imposes strict technical standards but also emphasises algorithmic transparency, continuous system performance monitoring, and the interpretability of decision-making mechanisms.¹⁰⁹ This places a broader duty on manufacturers, extending beyond formal compliance to proactive professional conduct and adaptation to scientific developments.¹¹⁰ For instance, if damage results from a fault in an algorithm trained on unbalanced data, the manufacturer bears the burden of proving that it took all necessary measures to prevent such errors or mitigate their effects.

The fundamental dilemma lies in delineating the boundary between harm accepted as part of “residual risks” and unjustified failures in design or operation.¹¹¹ Accordingly, proposals have emerged to assess liability based on comparing overall system performance against a “safe model” standard; if harmful error rates exceed a certain threshold, the design is considered defective, even if individual errors may be technically justified.¹¹² However, this approach faces technical challenges related to real-time monitoring, as well as political and ethical considerations involving acceptance of any potential harm to human life.¹¹³ Considering this, there is a pressing need to develop an integrated legal framework that connects technical compliance, professional responsibility, and ethical caution — affirming that justice cannot rest solely on textual adherence but requires a deeper perspective aligned with the complexities and rapid evolution of generative AI.

7 THE FUTURE OF CRIMINAL LIABILITY FOR ALGORITHMIC HARM

Traditional ideas of criminal liability are being called into question by the emergence of generative artificial intelligence, particularly as it permeates business operations. It is becoming more difficult to assign blame for damage brought on by autonomous AI decisions. To predict the future of criminal law amid profound technological change, emerging research aims to reconstruct legal frameworks suited to corporate AI.

7.1. Rebuilding Generative AI's Criminal Liability

Despite notable progress in developing frameworks for criminal liability in the age of artificial intelligence, the legal pathway to holding producers of AI systems, such as ChatGPT, criminally responsible remains complex and ambiguous. Features of these systems—from unpredictability to the opacity of algorithmic decision-making.¹¹⁴ This creates an epistemic gap between human action and harmful outcome, undermining the establishment of a clear causal link and weakening the logic of blame based on negligence or fault.¹¹⁵ This phenomenon has been described in legal literature as the "control gap crisis," which criminal law faces in a technological risk society, where traditional models centred on criminal conduct and free will fail to accommodate the emerging complexities of algorithmic actions. For instance, a decision by an autonomous vehicle’s algorithm may cause an accident without direct human intervention, raising profound questions about criminal liability.

Traditional criminal liability, which is predicated on intent or negligence, struggles to handle algorithmic decisions. Modern approaches emphasise duty of care violations when AI systems endanger fundamental rights, such as psychological or physical safety, placing a higher priority on preventive liability. Laws such as Article 12 of the UAE consumer protection law No (15) of 2020, which penalise dangerous products but still partially exempts software-only AI, reflect this trend. Thus, liability shifts from fault and intent to adherence to preventive and precautionary obligations.

The particularity of algorithmic harm calls for legislative renewal to keep pace with technological changes by treating AI as a legal entity with tangible effects—even absent physical embodiment. The European AI Act represents a significant step forward, imposing effective, balanced, and deterrent sanctions on violators of AI system regulations.¹¹⁶ The concept of a "legal warning" model also emerges, criminalising behaviours such as neglecting security measures or failing to update systems, with a combined regime of administrative oversight and criminal penalties to manage risks.¹¹⁷ Within this framework, authorities may be empowered to issue legal orders mandating additional testing, compulsory updates, or partial system suspension, with violations triggering criminal liability—thereby strengthening societal legal protection.¹¹⁸

The nature of algorithmic harm demands establishing a novel legal concept that redefines the relationship between algorithmic acts and legal efficacy, ensuring harm does not escape accountability under the guise of technology and innovation.¹¹⁹ Accordingly,

enhancing algorithmic transparency by obliging AI producers, such as ChatGPT, to regularly disclose potential risks supports a "democratisation of risk management" approach.¹²⁰ This recognises the right of society and legislators to be informed of the potential impacts of technology permeating all aspects of life. Consequently, it becomes essential to combine principles of strict or risk-based liability with advanced preventive legal mechanisms to balance the promotion of innovation with the protection of fundamental legal values such as safety and dignity, thereby achieving legal stability and criminal justice in addressing AI-caused harms.

7.2. Future Studies: Corporate Criminal Law and AI Liability

Traditional ideas of criminal liability are coming under increasing pressure from the emergence of generative artificial intelligence, especially as it becomes increasingly ingrained in business settings.¹²¹ The complexity of determining who is responsible for damage brought about by autonomous AI decisions has led to new research aimed at rebuilding legal frameworks suitable for corporate AI.¹²² This changing course aims to anticipate the future of criminal law and guarantee its flexibility in response to the significant technological advancements influencing contemporary responsibility and governance.¹²³

Future studies will focus on identifying the parties accountable for damage brought about by generative AI systems in commercial settings.¹²⁴ It is expected that questions will arise about the extent to which traditional notions of corporate criminal liability, such as vicarious responsibility or failure to supervise, can address harms caused solely by AI algorithms. It might be necessary to develop new legal frameworks that enable courts to hold companies responsible for algorithmic decisions.¹²⁵

Future research is expected to focus heavily on evaluating businesses' preventive responsibilities when using artificial intelligence. This could mean examining how much corporate governance, risk monitoring, and legal compliance are required of businesses, and how failing to comply could be interpreted as criminal negligence or complicity under corporate criminal law.¹²⁶

Comparative analyses of various legal systems are also likely to become increasingly important, offering insights into how conventional legal doctrines interact with the novel

challenges posed by AI-related harms. Such studies may support lawmakers and prosecutors in creating unified regulatory standards and future guidelines to control corporate risks related to artificial intelligence.

The notion of legal capacity for systems like ChatGPT involves treating AI as a potential legal entity with unique rights and responsibilities.¹²⁷ At present, global criminal law views AI as a tool for holding humans accountable, even as scholars point out that the existing legal system is not yet fully equipped to address accountability and regulatory questions raised by AI. An industrial or electronic form of legal capacity has been proposed as a potential model for structuring this recognition. Moreover, ChatGPT or AI systems would be punished by operational or technical limitations, such as halting operations or requiring updates, that are intended to prevent harm and ensure compliance, rather than the typical human-centred sanctions.

8 CONCLUSIONS AND RECOMMENDATIONS

The examination of criminal liability for artificial intelligence technologies, especially conversational systems like ChatGPT, reveals profoundly beneficial aspects that transcend the theoretical framework. Legislators are forced to reconsider traditional ideas of criminal liability as well as the actus reus (the act) and mens rea (intent) components of crime, considering the increasingly intertwined behaviour of humans and machines. It also provides legal professionals with new analytical tools to understand technological risks and assess their obligations in complex digital environments. Therefore, this study contributes to the development of future criminal policy toward a more flexible and equitable system in the age of artificial intelligence, while also improving scholarly discourse.

Fault-based liability is undermined by the difficulty of assigning criminal liability for algorithmic harm, as AI decisions often lack the traditional human intent required for criminal liability. Fundamental ideas in criminal law must be examined in light of the complexity and loss of control over technology. Legal standards that integrate technological tools to enhance preventive oversight and transparency are necessary to ensure accountability for harm caused by AI while balancing innovation and the defence of fundamental rights.

ChatGPT's legal capability could be electronically acknowledged, and operational or technical safeguards could ensure compliance and prevent harm, thereby ensuring accountability. Establishing a direct causal link between algorithmic actions and harmful outcomes is complicated by the opaque, complex nature of AI systems' decision-making processes. Traditional criminal law struggles to effectively address unpredictable, uncontrollable technological risks. There is a growing need to shift from fault-based

liability to a preventive liability model that emphasises the failure to implement adequate protective measures. Existing legal frameworks, including consumer protection laws, are inadequate for addressing the unique challenges posed by AI, especially for software that lacks physical integration. Transparency and regulatory oversight are critical to managing AI-related risks, with regular disclosure and mandatory updates serving as key tools to mitigate potential harm.

It is recommended that a legal framework be established that recognises AI's electronic capabilities and outlines operational or technical procedures as accountability mechanisms. Legal frameworks should be modernised to recognise AI as a distinct legal entity, capable of bearing responsibilities regardless of its physical form, with clear duties imposed on developers and users. A preventive liability approach should be adopted, criminalising negligence in failing to ensure safety measures and system updates, with deterrent penalties. AI developers must be required to provide regular transparency reports about potential risks, supporting broader societal and legislative oversight mechanisms. Administrative authorities should be empowered to monitor AI systems, enforce mandatory testing and updates, and impose partial shutdowns, when necessary, with criminal sanctions for noncompliance to ensure effective protection. REFERENCES

FOOTNOTES

REFERENCES

AUTHORS INFORMATION

RIGHTS AND PERMISSIONS

EDITORS

ABOUT THIS ARTICLE

DETAILS FOR PUBLICATION

AI DISCLOSURE STATEMENT

АНОТАЦІЯ