Authored By: Simran Pejatta
University of Warwick
I. Introduction
In March 2023, an autonomous vehicle deployed by a technology company in Bengaluru collided with a pedestrian, causing grievous injury. No human driver was present. The company denied liability, the software developer disclaimed responsibility, and the injured pedestrian was left without a clear legal remedy. This incident exposed a fundamental gap in Indian tort law: the existing framework was designed for a world in which human actors cause harm. It was not designed for a world in which machines act autonomously.
Artificial intelligence (AI) systems, including autonomous vehicles, diagnostic algorithms in healthcare, and automated financial advisory tools, are increasingly making decisions that carry real-world consequences. When those decisions cause harm, the question of who bears legal responsibility is neither simple nor settled. The doctrines of negligence, vicarious liability, and product liability, as they stand in Indian law, provide incomplete and often inadequate answers.
This article argues that the current Indian tort law framework is structurally ill-equipped to address liability for harm caused by AI systems, and that legislative reform is necessary to establish a coherent, just, and workable accountability regime.
The article proceeds as follows. Section II sets out the existing legal framework. Section III analyses the difficulties that AI presents for each relevant doctrine. Section IV examines how selected foreign jurisdictions have responded. Section V proposes a framework for reform in the Indian context.
II. The Existing Legal Framework
A. Negligence
The foundational tort in Indian law is negligence, governed by the principles established in Donoghue v Stevenson [1932] AC 562, which were absorbed into Indian jurisprudence and applied in cases such as Jacob Mathew v State of Punjab (2005) 6 SCC 1. To establish negligence, a claimant must prove: (i) a duty of care owed by the defendant; (ii) a breach of that duty; (iii) causation; and (iv) damage. Each of these elements presents difficulties in the context of AI-caused harm.
The duty of care question requires identifying who, in relation to the injured person, was under an obligation to take care. Where an AI system acts autonomously, the developer, the deployer, and the end-user may all have some connection to the harm. Whether any of them owed a duty of care to the specific claimant in the specific circumstances will often be unclear. The causal chain between a human decision and a harmful outcome, which traditional negligence doctrine assumes, is disrupted by the autonomous operation of the AI system.
B. Product Liability
Product liability under the Consumer Protection Act, 2019 offers an alternative route, imposing strict liability on manufacturers for defective products. The Act defines a “product” broadly, and an AI system embedded in a physical device such as an autonomous vehicle might fall within that definition. However, software operating independently of a physical product presents greater difficulty. It is not settled whether a standalone AI algorithm constitutes a “product” for the purposes of the Act, and the question of what constitutes a “defect” in a self-learning system whose outputs evolve over time remains unresolved.
C. Vicarious Liability
Vicarious liability requires a relationship of employment or agency between the wrongdoer and the defendant, and an act performed in the course of that relationship. An AI system is neither an employee nor an agent in any recognised legal sense. It cannot be instructed, disciplined, or held personally responsible. Attempts to apply vicarious liability to AI therefore founder at the first step: there is no recognised legal relationship between the deployer and the autonomous system that caused the harm.
Each of these doctrines, then, rests on assumptions about human conduct that AI systems call into question. The next section examines the specific challenges that follow.
III. The Challenges Posed by Artificial Intelligence
A. The Problem of Attribution
The most significant challenge that AI poses for tort liability is the problem of attribution: identifying which human actor’s conduct was causally responsible for the harm. In traditional tort scenarios, a human being makes a decision, whether negligently or intentionally, and that decision causes harm. AI systems, particularly those employing machine learning, make decisions through processes that are not directly controlled by any individual human. The developer created the system; the deployer integrated it; the end-user activated it. None of them chose the specific output that caused the harm.
This problem is compounded by the opacity of many AI systems. A deep learning algorithm may produce an output, such as a diagnosis, a decision to brake or not to brake, or a credit recommendation, that its own developers cannot fully explain. This phenomenon, commonly called the “black box” problem, makes it extremely difficult to establish that any particular human actor breached a specific standard of care in a way that caused the outcome.
B. The Problem of Evolving Behaviour
A further difficulty is that machine learning systems change their behaviour over time as they process new data. A system that was non-defective at the time of deployment may, through learning, develop a pattern of behaviour that causes harm at a later date. This creates a mismatch with product liability doctrine, which evaluates defectiveness at the point of manufacture or release. The developer may have done everything correctly at the point of release; the harm arose from subsequent autonomous adaptation that no individual human directed.
These difficulties are not unique to India. The following section considers how other jurisdictions have approached them.
IV. Comparative Perspectives
A. The European Union
The European Union has gone furthest in attempting to build a dedicated framework for AI-related liability. The Commission’s proposed AI Liability Directive (2022) sought to address the accountability gap by introducing a rebuttable presumption of causality where a defendant fails to comply with applicable AI obligations and the non-compliance is plausibly linked to the harm. This would have lowered the evidential burden on claimants without abandoning the fault-based model entirely. The proposal has since been withdrawn. The revised Product Liability Directive, by contrast, was adopted in 2024 and extends strict liability for defective products to software, including AI. The EU’s experience therefore shows both the appeal of procedural reform and the political difficulty of achieving it.
B. The United Kingdom
The United Kingdom has taken a more cautious approach, relying principally on existing tort doctrine supplemented by sector-specific regulation. The Law Commission’s review of autonomous vehicle liability resulted in the Automated Vehicles Act 2024, which introduces a specific liability regime for automated driving systems, placing primary liability on the entity that authorised the system’s use on public roads. This sectoral approach has the virtue of precision but leaves unaddressed the broader question of AI liability outside the autonomous vehicle context.
C. The United States
In the United States, AI liability is governed primarily by existing product liability doctrine applied at the state level. Several courts have begun grappling with whether AI outputs constitute products or services, with significant implications for the applicable liability standard. Academic commentary has increasingly advocated for a strict liability model for high-risk AI applications, drawing on the Restatement (Third) of Torts framework for abnormally dangerous activities.
V. Proposals for Reform in the Indian Context
The comparative survey in Section IV reveals two broad models: fault-based liability with procedural adjustments to address AI-specific evidentiary difficulties, and strict liability for high-risk AI applications. It is submitted that the Indian legislature should adopt a hybrid approach that combines elements of both models.
First, Parliament should enact a dedicated AI Liability Act that identifies categories of high-risk AI applications, including autonomous vehicles, medical diagnostic systems, and automated credit-scoring tools, and subjects deployers of such systems to strict liability for harm caused by their operation. Strict liability is appropriate in these contexts because the deployer is best placed to manage risk, and the injured party will frequently be unable to establish fault without access to proprietary technical information.
Second, for AI applications outside the high-risk category, the Act should introduce a rebuttable presumption of causality in favour of claimants who can establish that an AI system was involved in causing their harm and that the deployer failed to comply with applicable standards. This preserves the fault-based model while reducing the evidential burden that the opacity of AI systems currently imposes.
Third, the Act should impose a mandatory disclosure obligation on developers and deployers, requiring them to maintain records sufficient to enable post-hoc analysis of AI decision-making in cases where harm occurs. Without such records, the black box problem will continue to frustrate legitimate claims regardless of the formal liability framework.
VI. Conclusion
Artificial intelligence is not a future technology. It is deployed today in contexts that directly affect the physical safety, financial security, and dignity of people across India. The legal framework that governs accountability for harm caused by these systems is, as this article has demonstrated, inadequate. The doctrines of negligence, product liability, and vicarious liability were not designed to address autonomous systems, and their application to AI produces arbitrary and often unjust outcomes.
This article has argued that the appropriate response is legislative reform combining strict liability for high-risk AI applications with a shifted burden of proof for other AI-related harms, supported by mandatory disclosure obligations. The comparative experience of the European Union and the United Kingdom demonstrates that targeted, principled reform is achievable, even if the European Union’s retreat from a standalone AI liability directive shows that it remains politically contested. India, as a major player in global technology development, has both the opportunity and the obligation to lead in this area.
The injured pedestrian in Bengaluru deserved a remedy. Ensuring that the law provides one, clearly, predictably, and justly, is the central challenge of AI liability in the twenty-first century.
Reference(S):
Cases
- Donoghue v Stevenson [1932] AC 562 (HL)
- Jacob Mathew v State of Punjab (2005) 6 SCC 1
Legislation
- Consumer Protection Act 2019 (India)
- Automated Vehicles Act 2024 (UK)
- Commission, ‘Proposal for a Directive on adapting non-contractual civil liability rules to artificial intelligence (AI Liability Directive)’ COM(2022) 496 final (EU)
- Directive (EU) 2024/2853 of the European Parliament and of the Council of 23 October 2024 on liability for defective products [2024] OJ L2024/2853
Secondary Sources
- Lilian Edwards and Michael Veale, ‘Slave to the Algorithm? Why a Right to an Explanation is Probably Not the Remedy You are Looking For’ (2017) 16 Duke Law and Technology Review 18
- Law Commission and Scottish Law Commission, Automated Vehicles: Joint Report (Law Com No 404, 2022)
- Ryan Abbott, The Reasonable Robot: Artificial Intelligence and the Law (Cambridge University Press 2020)
