The Hallucination Tort: Liability in Generative AI

I. The Erosion of the Safe Harbor

For three decades, the internet has operated under the 'Safe Harbor' protection of Section 79 of India's IT Act (and Section 230 in the US). These statutes protected intermediaries from liability for content posted by third parties. However, Generative AI models are not intermediaries. They do not 'host' content; they 'generate' it. When an AI produces a defamatory statement, the developer is not merely a host—they are the architect of a system that actively creates original, albeit synthetic, harmful content. We are currently witnessing the collapse of the Safe Harbor for AI labs, as courts begin to recognize that a model's weights are not a neutral platform but a biased, generative engine.

Advocate R Khurana argues that the 'Safe Harbor' was never intended to protect autonomous agents. If an entity builds a machine that is known to periodically produce libelous falsehoods, and releases that machine to the public without sufficient guardrails, that entity is engaging in negligence. The legal standard must shift from 'Intermediary Immunity' to 'Product Liability'. Unlike a social media post, which originates from a human user, an AI output originates from the statistical patterns of the model itself. Therefore, the model's creator must be held to the same standards as any other publisher of information.

II. The Duty of Factual Fidelity

Does a developer owe a duty of care to the individuals contained within its training data? We propose the 'Duty of Factual Fidelity'. In the training of an LLM, the model consumes billions of tokens representing human identities. When the model "hallucinates"—producing a plausible-sounding but factually incorrect claim about an individual—it is a breach of this duty. Liability should not depend on whether the developer *intended* the lie, but whether they *permitted* the machine to speak without a verification mechanism.

"Liability is not a ghost in the machine; it is the responsibility of the architect who built the machine to lie so convincingly."

III. Stochastic Defamation and the Right to Reputation

Traditional defamation law requires 'publication' to a third party. In the case of a private chat with an AI, is the output 'published'? We argue that yes, the moment the model serves the content to any individual, publication has occurred. Furthermore, the harm of 'Stochastic Defamation' is unique: the AI can produce the same lie to millions of people simultaneously, or vary the lie slightly for each user, making traditional takedown orders nearly impossible to enforce. The reputation of an individual is a fundamental right under Article 21, and the "black box" nature of AI should not be allowed to act as a shield against the enforcement of this right.

The "Reasonable Person" standard in tort law must also be adapted for AI. A reasonable developer in 2024 should be aware of the "stochastic parrot" phenomenon and implement active truth-grounding and retrieval-augmented generation (RAG) to mitigate harm. Failure to do so constitutes a departure from the expected standard of care in the technology industry.

IV. Negligent Design and Algorithmic Bias

Hallucinations are often not random; they are biased. If a model is trained on data that contains systemic prejudice, its hallucinations will likely reflect and amplify those prejudices. This brings AI liability into the realm of Article 14 (Equality before the Law). If an AI hallucination leads to the denial of a loan or a job based on a synthetic falsehood, that is a state-sponsored act of discrimination (if the state is using the tool) or a massive civil tort (if a private entity is). The "Design Defect" argument from product liability law fits perfectly here: the model is defective because it is incapable of distinguishing truth from statistical probability.

V. Conclusion: Toward an AI Tort Code

The Hallucination Tort is the first step toward a broader framework of Algorithmic Tort Law. Advocate R Khurana continues to litigate at this frontier, ensuring that the march of progress does not leave the dignity of the individual in the dust of synthetic falsehoods. We need a comprehensive AI Tort Code that clearly defines the liability of developers, deployers, and users in this new generative economy.

The Chambers of Adv R Khurana are leading the charge in redefining the rules of engagement for the Artificial Intelligence age. Our goal is to ensure that the "intelligence" in AI is matched by the "accountability" of its creators.

When a generative AI system creates defamatory content, infringes copyright, or causes economic harm, who bears legal responsibility? The AI developer? The user who prompted the output? The platform hosting the AI? Or does the AI itself, as an autonomous agent, bear liability? These questions, once theoretical, have become urgently practical as generative AI proliferates across India. The Delhi High Court's ongoing ANI v. OpenAI case and multiple deepfake defamation suits signal that Indian courts are grappling with AI tort liability without clear legislative guidance. India's existing tort, contract, and consumer protection frameworks offer partial solutions but were not designed for autonomous AI systems, creating significant gaps in liability attribution and victim compensation.

I. The Liability Attribution Challenge

Traditional tort law requires establishing: duty of care, breach of that duty, causation, and damages. AI complicates each element. AI systems lack legal personhood and cannot be sued directly. They lack mens rea (mental intent), making it difficult to establish fault-based liability. The "black box" problem—the opacity of AI decision-making—makes proving causation challenging. Multiple parties (developers, deployers, users) contribute to AI outputs, fragmenting responsibility.

Indian law is exploring several liability models: Negligence—developers/deployers failed to exercise reasonable care in AI design, training, or deployment. Strict liability—liability without fault for inherently dangerous activities or defective products. Vicarious liability—holding employers/principals liable for AI actions within the scope of their business. Product liability—treating AI as a defective product under Consumer Protection Act, 2019.

II. Defamation by AI

Generative AI, particularly through deepfakes, poses unprecedented defamation risks. Indian defamation law (Sections 499-500 IPC, Section 356 Bharatiya Nyaya Sanhita 2023) requires publication of false statements harming reputation. AI-generated deepfakes can create convincing false videos of public figures making statements they never made. Recent cases—Arijit Singh v. Codible Ventures LLP, Aaradhya Bachchan v. Bollywood Time—highlight the urgency.

The Delhi High Court's landmark ruling in Ms. Kamya Buch v. Unknown (2024) established that platforms hosting AI-generated defamatory content can be held liable if they fail to remove it promptly after notification. However, determining who is the "publisher" for AI-generated content remains contested. Is it the AI developer, the user who prompted the generation, or the platform hosting it? Courts are developing a graduated approach: users who knowingly generate and distribute defamatory AI content bear primary liability; platforms lose safe harbor if they fail to act on complaints; developers may be liable if AI was designed or trained in ways that foreseeably enable defamation.

III. Copyright Infringement

The ANI v. OpenAI case (Delhi High Court, 2024) represents India's first major judicial examination of AI copyright liability. ANI alleges OpenAI trained models on copyrighted news content without authorization and that ChatGPT reproduces ANI content when prompted. The case raises fundamental questions: Does AI training on copyrighted material constitute infringement? Is it covered by fair dealing exceptions (Section 52, Copyright Act 1957)?

The Department for Promotion of Industry and Internal Trade (DPIIT) clarified in February 2024 that commercial AI training on copyrighted material requires authorization—fair dealing does not cover it. This creates potential liability for AI developers who trained models on copyrighted Indian content without licenses. For AI-generated outputs that reproduce copyrighted material, liability likely falls on both the user (who prompted the infringing output) and potentially the developer (if the AI was designed to enable such reproduction).

IV. IT Act and Intermediary Liability

Section 79 of the IT Act, 2000, provides safe harbor for intermediaries who don't initiate transmission, select receivers, or modify content. The IT Rules 2021 require platforms to label AI-generated content and work to curb deepfakes. However, the IT Act wasn't designed for AI's autonomous nature. Platforms hosting AI-generated content can lose safe harbor if they fail due diligence obligations: removing prohibited content within 24 hours of complaint, implementing technology to identify and remove CSAM and terrorism content, and labeling AI-generated content.

The proposed Digital India Act aims to replace the IT Act with more comprehensive provisions for algorithmic accountability and platform liability related to AI. Expected provisions include mandatory AI content labeling, algorithmic transparency requirements for high-risk AI systems, and clearer liability frameworks for AI-generated harms.

V. Consumer Protection Act 2019

The Consumer Protection Act, 2019, introduces product liability for defective goods or services, potentially extending to AI systems causing harm. Manufacturers and service providers may be liable for compensation if a defect is established, even without proving negligence. The Central Consumer Protection Authority (CCPA) can issue guidelines protecting consumers from unfair trade practices, including AI-facilitated false endorsements using celebrity likenesses.

A January 2026 amendment bill proposes broadening "unfair trade practices" to include algorithmic influence on consumer behavior, directly addressing AI manipulation concerns. However, applying the CPA to AI faces challenges: proving causation with opaque models, consumer courts declining complex technological cases, and determining whether AI is a "product" or "service."

VI. Practical Guidance

For AI developers: implement robust content filtering, maintain detailed training data records, provide clear user guidelines prohibiting harmful uses, establish rapid response mechanisms for complaints, and consider liability insurance. For AI users: understand terms of service and liability allocation, avoid generating content that could defame or infringe rights, verify AI outputs before publication, and maintain records of prompts and outputs. For platforms: implement AI content labeling, establish efficient complaint mechanisms, develop technology to detect deepfakes and prohibited content, and maintain clear liability disclaimers.

VII. Conclusion

Generative AI tort liability in India is evolving through judicial decisions, regulatory guidance, and legislative proposals. While existing frameworks provide partial recourse, comprehensive AI-specific legislation is needed. The Digital India Act, expected in 2026-2027, should establish clear liability frameworks, mandatory insurance requirements for high-risk AI, algorithmic transparency obligations, and victim compensation mechanisms. Until then, courts will continue developing AI tort law case-by-case, creating precedents that will shape India's AI governance for decades.

The Hallucination Tort

Advocate Ritik Khurana