1. Introduction

Artificial Intelligence (AI) has transformed healthcare, transportation, banking, insurance, education, manufacturing and other sectors. AI systems can analyse large volumes of information, make predictions, generate content and assist in decisions that previously required human judgment. However, when an AI system causes injury, financial loss, discrimination, property damage or reputational harm, a fundamental legal question arises:

Who should be held civilly liable when an AI system causes damage—the developer, manufacturer, deployer, user, or another party?

Traditional civil law generally assumes that harm results from a human act, an omission, a defective product or a breach of a legal duty. AI complicates this approach because its outputs may depend on complex algorithms, training data, third-party components and decisions made after deployment.

For example, consider the following situations:

A self-driving vehicle causes an accident because its object-detection system fails to identify a pedestrian.

An AI medical diagnostic system recommends an incorrect treatment, resulting in injury to a patient.

A bank uses an algorithm that unfairly rejects loan applications because of discriminatory patterns in its training data.

A generative AI system produces a defamatory statement about an individual.

An industrial robot damages property after receiving an unexpected instruction from an AI-based control system.

Each situation raises questions concerning negligence, causation, product defects, contractual obligations, compensation and the allocation of responsibility.

The central objective of civil liability law is to provide compensation to injured persons, allocate losses fairly, encourage reasonable precautions and prevent future harm. The challenge is to achieve these objectives without making developers automatically liable for every unexpected output or allowing organisations to avoid responsibility merely because an algorithm made the immediate decision.

2. Meaning and nature of civil liability in AI

2.1 Meaning of civil liability

Civil liability is the legal obligation of a person or organisation to compensate another person for harm caused by a legally recognised wrongful act, omission, breach of duty or defective product.

In the context of AI, civil liability refers to the legal responsibility of relevant human or corporate actors for damage associated with the development, supply, deployment, operation or use of an AI system.

AI itself is not ordinarily treated as a separate legal person capable of bearing conventional civil liability. In most existing legal systems, liability is attributed to the natural or legal persons involved in designing, manufacturing, supplying, controlling or using the system.

2.2 Essential elements of civil liability

The precise requirements depend on the applicable jurisdiction and cause of action. Nevertheless, the following elements are commonly relevant.

1. Duty of care or legal obligation

The defendant must owe a duty recognised by law, such as a manufacturer's duty to take reasonable care or a professional's duty to act with appropriate competence.

2. Breach of duty or defect

There must be a relevant failure to meet the applicable legal standard, unless the claim proceeds under a strict or other no-fault liability regime.

3. Causation

The claimant must establish the required connection between the defendant's conduct or the relevant defect and the damage.

4. Legally recognised damage

The claimant must establish an injury or loss that the applicable law permits them to recover, such as physical injury, property damage or qualifying financial loss.

These elements should not be applied mechanically to every claim. For example, strict product liability may not require proof that a manufacturer was negligent, while contractual liability may turn on breach of an agreed obligation.

2.3 Why AI creates special liability problems

AI systems create at least six significant legal difficulties.

Opacity: The internal reasoning of a complex model may be difficult to explain or reconstruct.

Autonomy: A system may generate an output without a person approving each individual decision.

Distributed responsibility: Developers, data suppliers, manufacturers, cloud providers and users may all contribute to the outcome.

Causation difficulties: Harm may result from a combination of defective design, inadequate data, poor integration and human supervision.

Evolving systems: A model may change through updates, fine-tuning or continued learning after it has been supplied.

Information asymmetry: The defendant may possess the logs, training records, testing documents and technical evidence necessary to prove what went wrong.

Consequently, ordinary civil liability principles remain important, but courts may need to adapt their application to AI-related disputes.

3. Principal legal approaches to AI liability

3.1 Negligence-based liability

Negligence is established when a person or organisation fails to exercise the level of care required by law and that failure causes legally recognised damage.

In AI cases, negligence may arise where a developer fails to test a model adequately, a hospital deploys a diagnostic tool without appropriate safeguards, or an employer relies on a recruitment algorithm without reasonable validation.

The claimant normally needs to establish the relevant duty, breach, causation and damage, subject to the rules of the jurisdiction.

Example: A hospital adopts an AI diagnostic tool despite known limitations, fails to train its clinicians and ignores warning signs. If these failures cause a patient to suffer injury, the hospital or responsible professionals may be liable under applicable negligence principles.

Advantage: Negligence encourages reasonable precautions and allows liability to reflect the conduct of the parties.

Limitation: It can be difficult for an injured person to prove precisely how a complex model failed or which participant's conduct caused the injury.

3.2 Strict product liability

Strict product liability can impose liability for a defective product without requiring the claimant to prove that the defendant was personally negligent. The claimant must still establish the applicable statutory or common-law requirements, including defect, causation and damage.

In an AI-enabled product, a defect might involve unsafe design, manufacturing errors, inadequate warnings or a failure to provide necessary safety instructions.

Example: An AI-enabled industrial machine contains a foreseeable control-system defect that causes serious injury despite the operator following the supplied instructions. Product liability may provide a route to compensation.

However, not every AI service or software output automatically qualifies as a defective product under every legal system. The applicable product definition and liability rules matter.

3.3 Vicarious liability

Vicarious liability may make an employer or another legally responsible principal liable for a tort committed by an employee or agent in the course of the relevant relationship.

For AI-related harm, this principle may apply where an employee uses an AI tool negligently as part of their work. It does not automatically mean that every organisation is liable for every output produced by a system it uses.

Example: An employee of a financial institution relies on an AI-generated assessment without performing mandatory verification, resulting in a wrongful action against a customer. The institution's potential liability depends on the employee's duties, the nature of the relationship and the applicable law.

3.4 Contractual liability

AI-related disputes can also arise from contracts between developers, vendors, businesses and customers.

Relevant issues include:

Failure to provide the promised level of accuracy or availability.

Breach of warranties or representations.

Failure to implement agreed security safeguards.

Breach of data-handling obligations.

Failure to meet contractual testing, maintenance or support requirements.

Contractual remedies depend on the agreement, applicable legislation and the loss proved. A disclaimer cannot necessarily exclude liability where mandatory law or public policy prevents it.

3.5 Absolute liability and hazardous activities

Some jurisdictions impose particularly stringent liability for harm associated with hazardous activities. In India, the Supreme Court developed the doctrine of absolute liability for enterprises engaged in hazardous or inherently dangerous activities.

This doctrine may be relevant when AI is integrated into a hazardous industrial operation, but it does not establish a universal rule that every AI developer is absolutely liable for all AI-related harm. The nature of the enterprise, the activity and the applicable legal principles must be examined.

3.6 Emerging risk-based approaches

A risk-based approach distinguishes systems according to the seriousness and likelihood of the harm they may cause. High-risk applications, such as certain medical, employment, transport or safety-related systems, may warrant stronger testing, documentation, oversight and insurance requirements.

Risk regulation and civil liability are related but distinct: a regulatory breach may help establish fault or provide evidence, but the consequences for a private damages claim depend on the relevant legal framework.

4. Comparative approaches to AI liability in different jurisdictions

Different legal systems approach AI-related harm through their existing rules of tort, contract, product liability and statutory regulation. Their differences concern who bears the burden of proof, whether negligence must be established, how responsibility is allocated and what remedies are available.

4.1 India: Traditional tort law, consumer protection and absolute liability

India does not have a single, comprehensive statute establishing a general civil liability regime for every form of AI-related harm. Claims must generally be assessed under existing legal principles and applicable legislation.

A. Negligence under tort law

Indian courts can apply established negligence principles where an AI-related injury results from an unreasonable failure to take care.

Potentially liable parties include:

Developers who fail to take reasonable precautions against foreseeable harm.

Manufacturers who supply unsafe AI-enabled products.

Hospitals and professionals who use AI systems without appropriate oversight.

Employers or businesses whose employees negligently use AI tools.

Organisations that fail to implement necessary security or validation measures.

The claimant must establish the elements required for the particular tort action. The fact that an AI system produced an incorrect output does not, by itself, establish that a particular defendant was negligent.

B. Consumer Protection Act, 2019

The Consumer Protection Act, 2019 provides relevant remedies in qualifying consumer disputes. Its product liability provisions may apply to harm caused by a defective product, subject to statutory definitions, requirements and defences.

AI-enabled devices, medical equipment and other tangible products may therefore raise product liability questions. A defective AI-based service may also raise consumer protection or service-deficiency issues, depending on the circumstances.

However, the Act should not be interpreted as automatically making every AI developer or service provider liable for all errors produced by its system.

C. Absolute liability

In M.C. Mehta v. Union of India (1987), the Supreme Court established the doctrine of absolute liability for enterprises carrying on hazardous or inherently dangerous activities.

The doctrine is relevant to AI where an AI system is integrated into a hazardous industrial activity, but its application depends on the nature of the activity and the relevant legal requirements. It is not a general rule imposing absolute liability on every AI company.

D. Constitutional and public law remedies

Where AI is used by public authorities, algorithmic decisions may also raise issues of equality, fairness, arbitrariness and procedural legality under the Constitution of India. A public law remedy may be available where the relevant constitutional or administrative requirements are met.

Such proceedings are distinct from an ordinary private damages action. A constitutional violation does not automatically establish every element of a tort claim.

Overall assessment: India's present approach relies substantially on established legal doctrines. These provide a foundation for AI-related claims, but disputes involving opaque models, complex causation and access to technical evidence may expose practical difficulties.

4.2 United States: Negligence, product liability and algorithmic accountability

The United States does not have one uniform national tort law for AI liability. State law plays a major role, and the applicable standards may differ between jurisdictions.

The principal approaches include negligence, strict product liability, breach of warranty, misrepresentation and statutory claims.

Negligence: A developer, manufacturer or deploying organisation may be liable if its conduct falls below the applicable standard of care and causes legally recognised harm.

Product liability: Depending on state law, claims may concern a design defect, manufacturing defect or inadequate warnings. Whether standalone software or a particular AI service qualifies as a product can be a contested question.

Algorithmic decision-making: AI systems used in employment, credit, insurance or criminal justice may also be subject to relevant anti-discrimination, consumer protection and sector-specific laws.

An important feature of the US approach is its attention to evidence of the defendant's conduct, the foreseeable risks, the causal relationship and the precise legal claim.

Its principal difficulty is that a claimant may have to overcome substantial evidentiary and procedural barriers, particularly when the relevant model and its operational records are controlled by the defendant.

4.3 United Kingdom: Negligence, product liability and established common-law principles

The United Kingdom generally approaches AI-related civil liability through existing common-law and statutory frameworks rather than a universal AI-specific compensation regime.

Relevant areas include:

Negligence and the duty of care.

Product liability legislation.

Contractual duties and warranties.

Data protection and privacy law.

Vicarious liability where its requirements are satisfied.

A court may examine whether a reasonable developer or operator would have anticipated the risk, whether appropriate safeguards were adopted, and whether the relevant failure caused the injury.

The UK's established negligence jurisprudence is especially important because it distinguishes the existence of a duty of care from breach, causation and remoteness of damage.

However, the law does not establish that a developer is automatically responsible whenever an AI system produces a harmful result.

4.4 European Union: Risk regulation and product liability

The European Union combines preventive regulation with national civil liability systems and EU legislation.

Two instruments are particularly important:

1. EU Artificial Intelligence Act

The AI Act establishes a risk-based regulatory framework. Depending on the system and its use, it imposes requirements concerning matters such as risk management, data governance, technical documentation, transparency, human oversight and conformity assessment.

These regulatory obligations can be relevant when determining whether an organisation acted appropriately. However, regulatory non-compliance and civil liability are distinct legal questions; compensation still depends on the applicable cause of action and legal requirements.

2. Revised Product Liability Directive

Directive (EU) 2024/2853 modernises EU product liability law and expressly addresses software, including AI software, within its product framework. It also responds to certain evidentiary difficulties associated with complex products.

Its practical application depends on the Directive's implementation and applicable transitional provisions.

The EU's proposed AI Liability Directive, presented in 2022, sought to make certain fault-based claims easier to prove through evidence disclosure and rebuttable presumptions concerning causation. That proposal was withdrawn in 2025 and should not be described as an enacted EU liability regime.

European Parliament

+2

The EU's approach therefore illustrates an important distinction: a jurisdiction can regulate AI extensively without having enacted a separate, comprehensive civil liability statute for every form of AI harm.

4.5 Comparative summary

The table describes broad tendencies rather than exhaustive statements of every state's or member state's law.

5. Important case laws relevant to civil and AI liability

Legal qualification: The following judgments are established precedents relevant to AI liability by analogy. They should not all be represented as direct decisions about AI-generated harm. Their value lies in the principles they establish for negligence, product defects, hazardous activities, professional care and algorithmic decision-making.

Case 1. Donoghue v. Stevenson 19321932 AC 562

Jurisdiction: United Kingdom | Principle: Duty of care and negligence

Facts: A consumer drank ginger beer from an opaque bottle containing the decomposed remains of a snail. She alleged that the manufacturer's failure to take proper care during production caused her illness.

Legal issue: Did the manufacturer owe a duty of care to the ultimate consumer, even though the consumer had no direct contract with the manufacturer?

Judgment and principle: The House of Lords recognised that a manufacturer could owe a duty of care to the ultimate consumer where the product was intended to reach the consumer without reasonable opportunity for intermediate inspection.

Application to AI liability: The principle is important because AI-related harm may occur without a direct contractual relationship between the injured person and the developer.

For example, a person injured by an AI-enabled medical device might seek to establish that the manufacturer failed to take reasonable care in designing, testing or supplying the device.

The case supports the general proposition that a manufacturer cannot necessarily avoid responsibility merely because the injured person purchased or encountered the product through an intermediary.

Significance: It provides a foundational basis for analysing the duty of care owed by AI developers and manufacturers. It does not, by itself, resolve the special issues of model opacity, autonomous outputs or the allocation of responsibility among multiple AI participants.

Case 2. Caparo Industries plc v. Dickman 19901990 2 AC 605

Jurisdiction: United Kingdom | Principle: Duty of care and foreseeability

Facts: Investors relied on audited company accounts when making a takeover-related investment decision. The case concerned whether the auditors owed the relevant duty to those investors for that purpose.

Legal issue: Under what circumstances does the law recognise a duty of care for economic loss arising from statements or professional conduct?

Judgment and principle: The House of Lords examined foreseeability, proximity and whether it was fair, just and reasonable to impose a duty of care in the circumstances.

These considerations form an important part of English negligence analysis, although the precise test depends on the legal context and subsequent developments in the law.

Application to AI liability: Suppose an AI company provides a predictive analytics system to a business. A third party suffers financial loss after relying on a prediction generated by that system.

The question is not simply whether the loss was foreseeable. The court must also consider the relationship between the parties, the purpose of the information, the circumstances of reliance and the applicable rules governing economic loss.

Significance: This case helps distinguish foreseeable harm from harm for which the defendant actually owes a legally recognised duty. It can therefore help define the limits of AI developers' liability to downstream users and third parties.

Case 3. Rylands v. Fletcher (1868) LR 3 HL 330

Jurisdiction: United Kingdom | Principle: Strict liability for certain hazardous escapes

Facts: The defendant arranged for a reservoir to be constructed on his land. Water escaped through old underground mine workings and flooded the claimant's mine.

Legal issue: Could the defendant be liable for damage caused by the escape of a dangerous thing brought onto the land for a non-natural use, even without proof of negligence?

Judgment and principle: The House of Lords established a rule of strict liability in the relevant circumstances, subject to recognised limitations and defences. The scope of this rule has subsequently been restricted and developed through case law.

Application to AI liability: Consider an AI-controlled industrial installation that releases a dangerous substance or causes hazardous machinery to operate unpredictably.

The case provides historical context for liability based on the nature of an activity or the escape of a dangerous substance rather than ordinary negligence alone.

However, the rule does not mean that an AI system is inherently a dangerous thing or that all software failures trigger strict liability. The requirements of the relevant doctrine must still be met.

Significance: It illustrates the policy argument that certain unusually dangerous activities may justify allocating the resulting risks to the enterprise conducting them.

Case 4. M.C. Mehta v. Union of India (1987) 1 SCC 395

Jurisdiction: India | Principle: Absolute liability for hazardous enterprises

Facts: The case arose from the leakage of oleum gas from the Shriram industrial complex in Delhi. The proceedings raised questions concerning the responsibility of an enterprise engaged in hazardous activities and the protection of people exposed to the resulting risks.

Legal issue: Should an enterprise engaged in hazardous or inherently dangerous activity be able to rely on the traditional exceptions associated with the rule in Rylands v. Fletcher?

Judgment and principle: The Supreme Court developed the doctrine of absolute liability for enterprises engaged in hazardous or inherently dangerous activities. Such an enterprise owes an absolute and non-delegable duty to the community and must compensate for harm resulting from the hazardous activity in accordance with the doctrine.

Application to AI liability: Imagine a hazardous industrial enterprise that uses AI to control chemical processing or other dangerous operations. If an AI-related failure causes harm, M.C. Mehta may provide relevant reasoning concerning the responsibility of the enterprise conducting the hazardous activity.

Nevertheless, the judgment does not automatically impose absolute liability on an ordinary software developer whose AI model is used in a hazardous setting. The nature of the defendant's activity and the requirements of Indian law remain essential.

Significance: This is particularly important in India because it demonstrates that liability principles can be adapted to technological and industrial risks when existing doctrines provide insufficient protection.

Case 5. Jacob Mathew v. State of Punjab (2005) 6 SCC 1

Jurisdiction: India | Principle: Standard of professional negligence

Facts: The case concerned allegations of medical negligence following the death of a patient. The Supreme Court examined the standards relevant to determining when a medical professional's conduct amounts to negligence.

Legal issue: What level of professional failure is required to establish negligence, and how should the law distinguish an adverse outcome from legally culpable conduct?

Judgment and principle: The Court discussed the standard of care applicable to medical professionals and the distinction between an unsuccessful outcome and negligence. It also addressed criminal medical negligence, so its criminal-law aspects must be distinguished from civil liability.

Application to AI liability: Suppose a doctor uses an AI diagnostic system that provides an incorrect result. The mere fact that the patient suffers harm does not necessarily establish negligence.

A civil court would need to examine the applicable professional standard, the information reasonably available to the doctor, the system's limitations, the degree of reliance, the availability of further tests and the conduct of other responsible parties.

The developer may also face a separate claim if the evidence supports a legally recognised defect or negligent act.

Significance: The case supports a careful, evidence-based approach to professional negligence. It cautions against equating every medical error or adverse outcome with liability while recognising that failure to meet the applicable standard of care may justify compensation.

Case 6. State v. Loomis, 2016 WI 68

Jurisdiction: United States, Wisconsin | Principle: Algorithmic risk assessment and procedural fairness

Facts: Eric Loomis challenged aspects of his criminal sentencing after the sentencing court considered a risk assessment produced using the COMPAS algorithm.

Legal issue: Could the use of a proprietary algorithmic risk assessment in sentencing raise due-process concerns, particularly when the defendant could not fully examine the underlying methodology?

Judgment and principle: The Wisconsin Supreme Court upheld the use of COMPAS in the circumstances before it, subject to cautions concerning its use and interpretation. The judgment recognised limitations associated with proprietary risk assessment, including concerns about transparency, accuracy and the appropriate weight given to algorithmic predictions.

Application to civil AI liability: Although this was a criminal sentencing case, its reasoning is relevant to algorithmic accountability more broadly.

For example, if a business relies on a proprietary AI model to determine credit eligibility, insurance risk or access to a service, the affected person may face difficulty understanding or challenging the decision.

The case illustrates why decision-makers should not treat an algorithmic score as infallible and why appropriate human review, disclosure and scrutiny can matter.

Significance and limitation: Loomis is not a civil damages precedent establishing liability for AI harm. Its relevance is comparative and procedural: it highlights legal concerns associated with opacity, proprietary systems and reliance on algorithmic assessments.

Case 7. Escola v. Coca-Cola Bottling Co., 24 Cal. 2d 453 (1944)

Jurisdiction: United States, California | Principle: Product safety and strict liability foundations

Facts: A bottle of Coca-Cola exploded while a waitress was handling it, causing injury. The litigation concerned responsibility for the allegedly defective bottle.

Legal issue: How should the law allocate the risk of injury caused by a product that allegedly becomes dangerous during ordinary handling?

Judgment and principle: The California Supreme Court addressed the evidentiary and negligence issues arising from the bottle's explosion. In a famous concurring opinion, Justice Roger Traynor argued for imposing strict liability on manufacturers for defective products, an argument that later helped shape American product liability law.

Application to AI liability: Consider a manufacturer that integrates AI software into a physical device. If the product injures a user because of a defect in its design or operation, product liability principles may be relevant even if the injured person cannot identify a careless act by a particular engineer.

Whether liability is established depends on the applicable product liability law, including the requirements for proving a defect and causation.

Significance: The case helps explain the development of strict product liability and the policy of placing certain product-related risks on manufacturers rather than requiring every injured consumer to reconstruct the precise act of negligence.

Case 8. Indian Medical Association v. V.P. Shantha (1995) 6 SCC 651

Jurisdiction: India | Principle: Medical services and consumer protection

Facts: The Supreme Court considered whether medical services fell within the scope of the Consumer Protection Act, 1986, and addressed the treatment of different categories of medical services under the statute.

Legal issue: In what circumstances can patients use consumer protection law to pursue remedies concerning medical services?

Judgment and principle: The Court held that medical services generally fell within the consumer protection framework where the statutory requirements were met, while distinguishing services rendered entirely free of charge and certain other arrangements.

Application to AI liability: Suppose a private hospital provides treatment using an AI-based diagnostic service, and the patient suffers harm because of a deficient service or an inadequately integrated medical system.

Depending on the facts, the patient may have a consumer claim against the hospital or another relevant provider, in addition to any available tort claim. Liability will depend on the statutory requirements, the relevant service relationship and the evidence of deficiency or defect.

Significance: The case demonstrates how established consumer protection mechanisms can remain relevant as healthcare delivery incorporates new technologies.

Case 9. Hedley Byrne & Co Ltd v. Heller & Partners Ltd 19641964 AC 465

Jurisdiction: United Kingdom | Principle: Negligent misstatement and economic loss

Facts: A business sought a bank reference concerning another company's financial standing and subsequently suffered financial loss. The dispute concerned responsibility for an inaccurate statement supplied in circumstances involving a disclaimer.

Legal issue: Can a person be liable for economic loss caused by a careless statement even when no physical injury or property damage occurs?

Judgment and principle: The House of Lords recognised that liability for negligent misstatement could arise where the relevant relationship and assumptions of responsibility supported a duty of care, although the disclaimer in the case prevented liability.

Application to AI liability: An AI analytics provider might supply a financial risk report or automated valuation that a customer reasonably relies upon. If the report is negligently generated or presented in circumstances giving rise to a duty, a negligent misstatement claim may be considered.

The result depends on the applicable rules governing reliance, responsibility, disclaimers and pure economic loss.

Significance: The case is useful for analysing AI systems that provide advice, predictions or reports rather than physical products.

6. Comparative significance of the case laws

The cases can be organised according to the legal question they help answer.

The comparative lesson is that existing legal principles already address many kinds of AI-related harm. The central difficulty is not necessarily the absence of all applicable law, but the challenge of applying established rules to technically complex systems and identifying the responsible party.

7. Who should be held liable when AI causes harm?

A major question in comparative civil law is how liability should be distributed among the participants in the AI lifecycle.

1. AI developer

Potentially liable for negligent design, inadequate testing, foreseeable model failures, or other legally actionable defects attributable to its conduct.

2. Manufacturer or product supplier

Potentially liable for a defective AI-enabled product, including relevant design, integration, manufacturing or warning defects.

3. Deploying organisation

May be responsible for choosing an unsuitable system, ignoring known limitations, failing to supervise its use, or deploying it without appropriate safeguards.

4. Professional or end user

May be liable where their own conduct breaches a legal duty, such as disregarding required checks or using the system negligently.

5. Data supplier or other technical participant

May bear responsibility where a legally relevant failure in data provision, integration, maintenance or another technical service contributes to the harm.

These categories are not mutually exclusive. Multiple parties may contribute to one injury, and the applicable law determines whether liability is joint, several, apportioned or otherwise allocated.

Illustrative example: An AI medical diagnostic error

Suppose an AI system fails to detect a serious disease, and a patient consequently suffers harm.

Developer: Did it fail to test the system adequately or address a known limitation?

Hospital: Did it deploy the system without suitable validation or clinical safeguards?

Doctor: Did the doctor unreasonably rely on the AI output or ignore contrary clinical evidence?

Manufacturer: Was there a defect in the integrated medical device?

Data provider: Did a relevant failure in the supplied data contribute to the error?

A court should examine the evidence against each party separately. The fact that the AI output was wrong does not establish that all five parties are liable.

8. Major challenges in establishing AI liability

8.1 The black-box problem

Some AI models produce outputs through processes that are difficult to explain in terms readily understood by a court.

This can make it difficult to establish:

Why the system produced a particular output.

Whether the output resulted from a defect or a foreseeable limitation.

Whether a developer or operator knew or should have known about the risk.

Whether the defendant's conduct caused the claimant's loss.

Technical opacity should not automatically establish liability, but neither should it automatically excuse a defendant from responsibility.

8.2 The problem of causation

Traditional negligence claims generally require a legally sufficient causal connection between the relevant breach and the injury.

With AI, the causal chain may involve training data, software design, model updates, user instructions, external sensors and human supervision.

A claimant might need technical evidence to distinguish among several possible causes. Courts must also determine whether the damage is sufficiently connected to the defendant's conduct under the applicable causation and remoteness rules.

8.3 Burden of proof

The claimant may not have access to source code, testing records, incident logs or model documentation.

This creates an imbalance between the injured person and the organisation controlling the system.

Potential legal responses include:

Court-ordered disclosure of relevant technical evidence.

Preservation of logs and system records.

Expert examination of the model and its deployment.

Rebuttable presumptions where legislation permits them.

Evidentiary consequences for unjustified failure to disclose required information.

These mechanisms must be applied under the relevant procedural and substantive law. They should not be confused with a universal rule reversing the burden of proof in all AI cases.

8.4 Foreseeability and autonomous behaviour

AI systems can produce unexpected results, but an unexpected result is not necessarily unforeseeable in the legal sense.

A court may consider the known limitations of the system, the intended use, the available warnings, the precautions reasonably expected and the circumstances of deployment.

A developer is not necessarily liable for every misuse or genuinely unforeseeable event. Equally, describing an output as autonomous does not automatically eliminate the responsibility of those who designed, supplied or deployed the system.

8.5 Compensation and insurance

AI-related injuries may involve several responsible parties and substantial losses. Determining who can pay compensation is therefore an important practical concern.

Possible policy measures include:

Mandatory insurance for specified high-risk applications.

Compensation funds for narrowly defined categories of serious harm.

Contractual allocation of risk between commercial participants.

Contribution claims between defendants after compensation has been paid.

Such measures may improve access to compensation, but they require appropriate legislative design and should not be treated as existing universal requirements.

9. How should civil liability law evolve to address AI?

A balanced framework should protect injured persons without making the development of useful AI systems legally impracticable.

9.1 Retain negligence as a general foundation

Negligence remains suitable for many disputes involving professional judgment, software development and organisational supervision. Courts can examine whether reasonable precautions were taken in light of the system's intended use and foreseeable risks.

9.2 Apply product liability where appropriate

Where AI forms part of a product within the relevant legal definition, product liability rules may provide a more effective remedy than a claim requiring proof of an individual engineer's negligence.

Legislation should clarify the treatment of software, updates, integrated products and defects arising from relevant post-sale changes.

9.3 Introduce proportionate evidentiary rules

Legislatures could establish targeted disclosure obligations and rebuttable presumptions for situations in which technical complexity creates an unusually severe evidentiary disadvantage.

Any such rules should have clear thresholds, appropriate safeguards and opportunities for defendants to rebut the presumption.

9.4 Strengthen documentation and auditability

Developers and deployers of high-risk AI systems should retain appropriate records concerning testing, system limitations, material updates and incidents.

Such documentation helps identify failures and provides evidence for courts, regulators and affected individuals.

9.5 Consider risk-based strict liability

Strict liability may be appropriate for carefully defined activities or products presenting substantial risks, particularly where those who introduce the risk are best placed to prevent or insure against it.

However, imposing strict liability on every AI developer for every harmful output could discourage beneficial innovation and allocate losses unfairly.

The appropriate scope of strict liability should therefore depend on the type of activity, the seriousness of the risk, the parties' control over it and the feasibility of prevention.

9.6 Protect fundamental rights

AI used in employment, healthcare, credit, education and public administration can affect equality, privacy, dignity and access to essential services.

Civil liability should operate alongside applicable anti-discrimination, privacy, consumer protection and constitutional safeguards. Not every violation necessarily results in a damages award, but the relevant legal remedies should be effective and accessible.

10. Critical evaluation of the comparative approaches

The Indian, American, British and European approaches illustrate different priorities.

India can draw upon negligence, consumer protection and its established doctrine of absolute liability for qualifying hazardous activities. Its principal challenge is ensuring that traditional remedies remain effective when technical evidence is difficult to obtain.

The United States offers a developed range of negligence and product liability doctrines. Its strength is the ability to assess claims under different established causes of action. Its challenges include variation among jurisdictions and the expense and complexity of technical litigation.

The United Kingdom relies on established common-law reasoning, supplemented by statutory product liability and other relevant legislation. Its strength is doctrinal flexibility, while its principal challenge is applying conventional concepts of duty, causation and economic loss to increasingly complex AI systems.

The European Union combines risk-based regulation with product liability legislation. This approach places considerable emphasis on preventing harm and addressing certain evidentiary difficulties. However, regulatory compliance and civil compensation remain distinct, and the withdrawal of the proposed AI Liability Directive means that the proposal's intended general procedural changes should not be treated as enacted law.

No single approach completely resolves all AI liability problems. A workable system requires a combination of prevention, fair attribution of responsibility, access to evidence and effective compensation.

11. Conclusion

Civil liability for artificial intelligence is best understood as an evolving application of established legal principles rather than a completely separate field of law.

The foundational cases—Donoghue v. Stevenson, Caparo v. Dickman, Rylands v. Fletcher, M.C. Mehta v. Union of India, Jacob Mathew v. State of Punjab, State v. Loomis, Escola v. Coca-Cola Bottling Co., Indian Medical Association v. V.P. Shantha and Hedley Byrne v. Heller—illustrate the doctrines that courts and legislators can draw upon when addressing AI-related harm.

Their significance differs: some concern negligence and product safety, others hazardous activities or consumer remedies, while Loomis provides a comparative illustration of concerns surrounding algorithmic decision-making rather than a direct civil liability precedent.

The central legal objective is to ensure that AI does not become a means of avoiding accountability. At the same time, liability must be based on applicable legal standards and sufficient evidence, rather than on the mere fact that an AI system produced an undesirable outcome.

A balanced framework should preserve negligence and product liability where appropriate, clarify responsibility among developers and deployers, improve access to technical evidence, and consider stricter liability for clearly defined high-risk applications. Such a framework can protect injured persons while supporting responsible technological innovation.

Academic note: This discussion is intended for legal study and comparative analysis. The precise outcome of an AI-related claim depends on the facts, the applicable jurisdiction, current legislation and the specific cause of action pleaded.