AI is characterized by numerous features that have contributed to its rapid spread and adoption across various fields. The most significant of which may be summarized as follows:

Autonomy: AI’s ability to make decisions and execute complex tasks without human intervention and operating decisions autonomously by determining appropriate means to achieve its objective.

Simulation: This characteristic constitutes the foundational basis of AI. The core concept of AI is to simulate the capabilities of the human mind in analysis, comprehension, and problem-solving. AI performs complex functions that typically require direct human intervention to process and execute.

Self-learning and adaptation: AI’s capacity to learn and derive lessons is among its most distinguishing features. AI can process information and big data using self-learning algorithms that formulate and test assumptions. The more information AI is exposed to, the more its performance evolves; its ability to detect patterns, infer results, adjust behavior, and autonomously analyze data develops through techniques such as Machine Learning and Deep Learning. (Abou Adel et al., 2024; Alhourani et al., 2026; Qiqieh et al., 2025; Adel, 2022; Abdulrahman, n.d.)

Prediction and decision-making: AI enables the processing of vast quantities of data at very high speed, empowering it to predict future outcomes or propose decisions based on statistical and algorithmic inferences.

Scalability and development: AI’s flexibility enables expanding its applications by increasing inputs and improving algorithms. This characteristic underscores AI’s dynamic and ever-evolving nature, with capabilities that are variable and constantly growing. ( Norvig, 2010).

Natural language processing (NLP): AI uses this capability to understand language details which allows it to produce correct answers that meet user requirements. The feature developed through the use of chatbots and voice-assistant applications which includes the virtual assistant ‘Siri’ created by Apple as its most important demonstration ( Nigam, 2024).

Errors and hallucinations of AI and their impact on civil liability

AI systems possess distinct attributes which enable them to operate independently while they learn new information and make future predictions. The implementation of artificial intelligence technology changes the established legal framework because its autonomous systems and self-learning capacities create legal challenges which extend beyond technical matters to affect fundamental tenets of contemporary legal systems ( Qutieshat et al., 2026). The most prominent of these issues include the following:

First - autonomy (agentic AI)

The autonomous capabilities of artificial intelligence systems create legal difficulties because the systems provide incorrect information which leads to incorrect final results and dangerous situations that result in harm to people and property. The key legal challenges arising from autonomy include: -

Legal liability: The autonomy of AI and its ability to perform tasks independently and without any human intervention raises a fundamental question: who bears the consequences of the system’s decisions, acts, or errors - the manufacturer, the owner, the operator, the programmer, or the user? This feature has driven some to adopt the hypothesis that a robot, as one of the most important AI applications, should acquire legal personality as an electronic or virtual person, such that it would be liable to compensate third parties ( Ben Othman, 2020). In our view, however, this hypothesis is not practically or legally workable; even if an artificial juristic personality could be created for intelligent machines and generative AI systems, such systems do not possess an independent Financial Liability from which compensation could be paid to the injured party.

Second - ‘AI hallucinations’

Hallucinations are one of the biggest issues facing AI systems, particularly ‘generative AI’ based on deep learning and self-learning. Such systems use mathematical algorithms and complicated internal working processes, making it difficult or impossible to anticipate all outputs or judgments under different conditions. When generative AI algorithms produce outputs that are not grounded in the data on which the system was trained, or when intelligent systems do not follow a recognized or explainable pattern of reasoning, ‘hallucination’ or ‘AI hallucinations’ may occur ^† .

To understand the causes of AI hallucinations, Professor Novotny explains that such hallucinations are a side effect arising from how generative AI operates: intelligent systems use an algorithm to generate outputs by sampling from probability distributions built from large sets of examples. In essence, that algorithm relies on a form of conjecture in arriving at outputs and results; accordingly, it may also produce incorrect statements or acts ( Nasr, 2025).

The reasons underlying AI hallucinations can be summarized as follows ( Coursera, 2025):

Limited training data: Biases in the inputs used to train AI models can lead them to produce biased and inaccurate hallucinations presented as if they were reliable facts.

Model complexity: If an AI model is so complex that it lacks constraints limiting the type and quality of outputs it can produce, it may generate outputs that are based primarily on conjecture.

Data poisoning: Data poisoning occurs when internet hackers inject incorrect, misleading, or biased data into the training datasets of an AI model, which may create a security issues or lead to a cyberattack.

Overfitting: An AI model may successfully predict training data but fail to generalize to fresh data. If an AI system is trained to recognize humans and given many images of people, including groups of people standing next to streetlights, the model may incorrectly learn that streetlights are humans and begin to identify them as people. An experimental test conducted by Stanford University in 2024, involving leading legal-AI platforms across over 200 legal queries, revealed that the +LexisNexis Lexis platform reached a hallucination rate of 17% of cases, while the hallucination rates in Westlaw AI-Assisted Research was 33%. The ‘Ask Practical Law AI’ platform provided an accurate responses to queries in only 18% of cases ( POLANI, 2025).

In short, we consider that the difficulty of predicting AI system outputs due to AI hallucinations is not merely a technical problem; rather, it is a legal problem connected to the rules of civil liability. For this reason, most legal systems tend to impose restrictions on AI outputs and decisions whenever they entail legal or social effects.

Third - technical errors

The occurrence of technical errors in AI systems is attributed to one of the following causes: •

Errors in data or input.

Beyond the technical domain, such errors trigger multiple legal issues, the most important include: the liability issue of identifying the actor responsible for the error; the transparency issue of explaining the automated decision produced by AI on the basis of that error; and the security issue of preventing intrusion and manipulation that may arise from any vulnerabilities emerging from such technical errors ^‡ .

Given the legal issues associated with these technical errors, modern legislations - foremost in the UAE - has embedded two general principles governing AI use: (i) the principle of safety, which aims to ensure that all AI systems comply with the highest safety standards and encourages the modification or removal of high-risk systems; and (ii) the principle of transparency, which calls for a clear understanding of AI and how systems operate and make decisions, thereby contributing to building trust and enhancing responsibility and accountability in the use of these technologies.

The harms addressed by civil liability rules differ from those arising from the use of AI systems, as the latter possess a distinct legal character. This creates a point of difficulty in light of the civil liability rules regulated by the UAE Civil Transactions Law ( UAE Civil Transactions Law, Federal Decree-Law No. (25) of 2025), which attributes damage and fault to a direct human act or to the custody (guardianship) of specific tangible things ( Article 246 and Article 271 of Federal Decree-Law No. (25) of 2025). Today, however, we face harms of a different nature arising from AI outputs, harms that are non-material and intertwined with autonomous capabilities. This reveals a new type of harm that is difficult to subject to the traditional standards for proving fault ( Faqih, 2024), or even identifying its source, especially given the overlap among the roles of the developer, operator, and user ( Surden, 2019).

Comparative trends, especially in Europe, are re-examining civil-liability concepts by expanding strict (objective) liability and linking it to risks associated with use of advanced technologies ^§ , as reflected in legislative proposals regulating AI and liability for its products (European Commission, White Paper on Artificial Intelligence: A European Approach to Excellence and Trust ., 2020). These proposals are characterized by taking into account the specificity of digital harm and the need to develop effective legal mechanisms that protect the injured party without stifling technological innovation.

The legal system needs to establish civil liability rules which protect injured parties through the identification of persons who bear responsibility for harmful acts through both contract law and tort law. Most legislation does not recognize an independent legal personality for technical systems. A smart system operating under a contractual agreement creates contractual liability when the debtor fails to fulfill his responsibilities which require him to either produce specific outcomes or show proper safety procedures or necessary information disclosure. When harm happens beyond the contractual connection, tort liability emerges on the basis of fault or liability for things;the AI system operator assumes the role of system custodian because he owns control over the system. The custodian becomes responsible for compensating the damaged party when he fails to inform about system dangers or conceals technical details from the public. Many legal scholars view AI systems as harmful and require tougher liability regulations to balance innovation with injury prevention.

Most jurisdictions do not consider AI systems as having autonomous legal personality thus civil-liability standards require attribution of the harmful action to an actual person or legal entity for contractual and tort liability. In a contractual connection (e.g., an AI system operating agreement) contractual liability exists when someone proves that a party broke an obligation which included safety duties or responsibilities to inform about important details. When harm occurs outside the contractual relationship tort liability arises on the basis of fault or under the UAE Civil Transactions Law on liability for things and equipment under a person’s control custodianship so the operator may be treated as the custodian when control and supervision are established. Thus, failing to warn of system dangers or hiding technical information from consumers may result in responsibility and compensation. Many legal scholars consider AI systems “dangerous things” which require tougher liability regulations to balance innovation with victim protection.

To address these harms, the European Union’s AI law adopted the idea of classifying the risks arising from AI uses into four categories: unacceptable risks, high risks, limited risks, and minimal (or acceptable) risks. This classification yields the following:

Unacceptable risk: AI systems that pose unacceptable risks, such as the right to dignity, education, work, privacy, equality, health, safety, and human freedom, are often banned because their severity outweighs their potential benefit. For instance, a school may utilize an AI system to grade pupils depending on their behavior, depriving low-scoring students of one or more of those fundamental rights and vice versa. This is social scoring. This category is illegal and unethical since it violates freedom, privacy, and fundamental rights and discriminates against people.

High risk: While this category is dangerous, the EU AI Act did not ban its usage by AI systems. In sensitive sectors involving individuals’ rights, such systems are utilized to screen job candidates’ CVs and provide an automatic acceptance or rejection decision based on prior training. Bias or inaccuracy in the automated decision may reject a qualified person without explanation. The hazard of this issue is its influence on individual rights and unlawful applicant discrimination. Thus, the European legislature required human oversight to ensure openness and bias-free decision-making.

Limited risk: This category does not harm human rights, such as when a person is mislead by a company’s website into thinking a human replies to inquiries when they are actually connecting with an AI system. Given the low severity of this risk, the European legislature allowed its usage without control, but required users to be told that they were dealing with an AI system.

Minimal (or acceptable) risk: This type is utilized in everyday transactions and does not violate human rights. It is non-binding, thus violating it has no legal consequence, such as when an AI program proposes arranging images on a phone or fixing spelling problems when writing. In such circumstances, the program only assists the user; the user can accept or reject the system’s proposal without legal consequences.

The difficulty of proving fault within AI systems

The UAE Civil Transactions Law grounds tort liability on its three elements: the harmful act (fault), damage, and causation, as provided by Article 246 of the new UAE Civil Transactions Law No. 25 of 2025, which stipulates: “Any harm caused to another shall obligate the perpetrator, even if non-discerning, to compensate for the damage.” Liability is thus understood as arising from a harmful act attributable to an identified actor (most often a human being). This traditional conception faces a fundamental challenge when applied to AI systems, particularly given the non-recognition of legal personality for AI ^** . The inadequacy of the Civil Transactions Law in addressing AI errors and hallucinations becomes evident. Technical errors arising from the use of an AI system do not result from direct human behavior capable of being assessed by the “reasonable-person standard”; rather, it may result from complex interaction between AI algorithms and self-learning mechanisms. Consequently, it becomes difficult to determine whether the error stems from the programmer, the user, the system provider, or the system itself, leading to difficulty in identifying the legally responsible actor ( Alarie, Niblett, & Yoon, 2018; Surden, 2019). In this context, some scholars ( Al-Sanhuri, 2018; Al-Sahli, 2022) have grounded liability for harms generated by AI uses on liability for things, as provided in Article 271of Federal Decree-Law No. (25) of 2025 issuing the Civil Transactions Law, under the subsection titled “Liability of the Custodian of Things, Animals, and Buildings,” which provides: “Whoever has under his control things requiring special care to prevent their harm, or mechanical machines, shall be liable for the harm caused by such things or machines, except what cannot be avoided, without prejudice to any special provisions in this regard.” This is on the basis that AI systems fall within the concept of “things.” However, AI may not correspond to the strict notion of a “thing,” ( Al-Sahli, 2022; Marqas, 1995; Shannab, 2012) as it occupies an intermediate position: it rises above a thing but does not reach the status of a natural or juridical person.

This deficiency becomes even clearer when proving causation, which the law requires for liability. Article 255 of the Civil Transactions Law No. 25 of 2025 provides: “Compensation shall, in all cases, be assessed according to the damage suffered by the injured party and the profit lost, provided that this is a natural consequence of the harmful act.”Accordingly, liability is established only where the damage is a natural consequence of the harmful act. This requirement conflicts with the non-linear and the opacity of some AI systems, known as the “black box” problem, which has been previously explained, making it difficult for the injured party to prove that the damage arose as a direct result of the smart system’s decision. It may become impossible to establish the link between act and outcome, which may negate causation, or render legal causation presumed yet exceedingly complex, thereby practically depriving the injured party of protection ( European Commission, 2022; Pagallo, 2018).

It follows from the foregoing that the traditional civil liability rules, in their current formulation within the UAE Civil Transactions Law, as well as the rules set out in the Consumer Protection Law, are insufficient or ill-equipped to deal effectively with the risks arising from AI errors and hallucinations ( Issa, 2022). Where the Civil Transactions Law requires proof of fault and causation in their classical sense, this does not align with the technical specificity of AI systems or the nature of their outputs. Thus, despite the actual occurrence of damage, the injured party may be unable to obtain compensation because they cannot meet the burden of proof imposed by the general rules. This reveals a genuine legislative gap: rules originally designed for traditional human acts are being applied to technological risks of an entirely different nature, which may conflict with the principle of compensatory justice, one of the core foundations of civil liability, and ultimately leads to weakened legal protection for the injured party ( European Commission, 2020).

In view of this deficiency, doctrinal calls are increasing to reconsider the traditional concepts of fault and causation when dealing with AI systems, by adopting more flexible legal models, such as presumed liability or risk-based liability ( Khater et al., 2025). This direction is regarded as a means of achieving a necessary balance between encouraging technological innovation, on the one hand, and ensuring effective protection for the injured party, on the other, without undermining the general rules of the UAE Civil Transactions Law ( Imad et al., 2026).

There is also a trend toward adopting strict liability, grounded in risk theory, as a legal mechanism to address damages arising from AI systems. This theory is premised on assigning liability to the person who benefits from the hazardous activity they have created, without requiring proof of a harmful act (fault). It is rooted in the principle of “liability accompanies benefit”, meaning that whoever derives benefit from a given activity must bear the consequences of its risks. When applied to AI uses, this concept implies that entities operating intelligent systems or deriving benefit from them would bear responsibility for compensating resulting damages ( Pagallo, 2018; Surden, 2019). However, this theory encounters a significant difficulty: the multiplicity of actors involved in the design, operation, and distribution of AI systems, from the manufacturer to the programmer, then the developer, to the supplier, the authorized representative, or even the distributor. Identifying the producer of an AI-related system is therefore difficult and surrounded by ambiguity and complexity ( Faqih, 2024).

Given the European Commission’s conviction that traditional liability rules are insufficient to address AI risks, it adopted the idea of easing the injured party’s burden of proof with respect to AI liability, alongside establishing legal presumptions that assume causation where technical or organizational obligations imposed on smart-system operators are breached ( European Commission, 2022).

In light of the inadequacy of civil liability rules to confront the risks of AI systems, including errors and hallucinations, it has become necessary to seek a legal solution to protect society from these risks, by moving from total reliance on traditional civil liability rules to adopting an innovative, flexible, and comprehensive legal vision that ensures full protection and compensation for those dealing with AI systems, in a way that fulfills one of the most important objectives of the UAE Civil Transactions Law, which is to compensate for the damage and prevent harm to others ( El-Erian et al., 2026; Nggilu et al., 2025). This requires considering the establishment of a set of preventive legal obligations that achieve a balance between encouraging innovation and, at the same time, considering the more important requirement of protecting the injured party. This solution proposes a legal vision based on strengthening the duty of transparency combined with the duty of warning, to reduce the risks associated with the use of AI: “Towards a flexible legislative framework that combines preventive obligations and compensatory liability,” as follows:

Transparency obligation for self-operating systems

This study believes that AI-powered technical systems must be transparent, as transparency is one of the most significant legal requirements to safeguard the harmed party from errors or hallucinations. Transparency helps the injured party and judicial authorities understand how the system operates, its correctness, and its hazards as much as feasible. This requirement follows the UAE Civil Transactions Law’s principles of good faith and non-harm. Transparency lets the judge assess fault and deduce it from the totality of facts and indicia, as well as identify the level of automated behavior’s causation and damage. Since concealing crucial information that prevents evaluation of the intelligent system’s detrimental behavior is proof of error, developers or operators are accountable if such opacity causes actual injury to others ( European Commission,2022; Surden, 2019).

A Chinese court has ruled that a developer is not legally responsible for an AI “hallucination”, establishing a precedent for how such cases may be handled under Chinese law. The ruling, issued by the Hangzhou Internet Court, classifies AI-generated content as a service rather than a product in cases involving hallucinations. As a result, users must prove that a developer was at fault in the content-generation process and that the error caused actual harm. The court said AI-generated content generally does not constitute high-risk activity and that developers have limited ability to control AI responses. Imposing strict liability, it said, could hinder technological innovation ( Qitong, 2026).

The duty to warn of hallucinations as an independent legal obligation

UAE legal doctrine and case law require AI system operators to warn of AI-generated hallucinations and output mistakes. AI system operators must warn of the likelihood of hallucinations or inaccuracies in its outputs as part of UAE law’s responsibility of disclosure and warning. Activities with novel dangers, like AI systems, increase its importance. If the developer or operator knows of deceptive or hazardous outcomes, failing to notify users is negligence and civil culpability. UAE courts consider failure of duty to warn a defect that harms others. In sensitive sectors, where technical trust alone is insufficient, unambiguous warnings about the system’s limits and risks are crucial ( Al-Rai et al., 2026; Pagallo, 2018).

Compulsory insurance for AI systems to compensate the injured party

Among the elements of the proposed legal vision is the proposal to impose a compulsory insurance system on the use of any high-risk AI systems, to ensure that the injured party obtains compensation without requiring them to bear the burden of proving errors and hallucinations of this complex technology. This proposal is consistent with the logic of guarantee under the UAE Civil Transactions Law, particularly in relation to activities involving special risks, as compulsory insurance contributes to a fair distribution of risk among multiple parties. Developers and users may purchase insurance against damages caused by AI, and this system is considered, in the opinion of jurisprudence, an appropriate solution to cover unexpected damages caused by AI ( Fadel, 2023).

The idea of establishing compensation funds emerged in several fields, such as compensating victims of traffic accidents where the perpetrator is unknown, and other accidents. Therefore, this idea emerged in the context of compensating persons harmed by AI errors and hallucinations, as a complementary solution to the current civil liability system, especially in cases where identifying the responsible person is difficult or where proving fault and causation is not feasible. Accordingly, compensation funds may be a mechanism that the legislator can adopt in exceptional circumstances, or last resort, to redress harm and compensate injured in a manner that achieves social justice and prevents leaving the injured party without effective protection. Such funds may be financed by developers and users of AI systems.

This view highlights the need to strengthen the civil judge’s role in assessing technical error and extending larger jurisdiction in presuming error or causation if damage is demonstrated. Since AI systems cannot be subjected to the fault standard used to human conduct, expanding the judge’s ability to determine technical error is one of the best ways to address their complexity. Thus, jurisprudence and the judiciary strengthen the judge’s ability to infer fault from facts and indicia, allowing the court to conclude fault or causation if damage is linked to a high-risk technical activity. Given the insufficiency of civil liability rules and the growing interaction with AI systems, this strategy may be the current solution until new legislation regulates AI-related injuries ( Surden, 2019).

In light of the foregoing, and with the increasing harms caused by AI, the question arises as to whether the current reliance on the general rules of civil liability along with corporate policies and AI ethics charters is sufficient to address these problems. The jurisprudence has been echoed between supporters and opponents, while voices calling for the need for special legislation regulating these matters have grown louder. The European experience has clearly confirmed that effective protection for the injured party is achieved only through binding legislative regulation that complements the jurisprudential and judicial roles. Accordingly, society awaits the issuance of a special and independent UAE legislation on the liability of AI, based on both prevention and compensation, in line with the country’s directions toward supporting responsible innovation while protecting individuals’ rights at the same time.

1.

Calling for developing a specific UAE legislative framework for AI, or introducing qualitative amendments to the general rules, taking into account the technical nature and newly emerging risks of these systems.