Copyright Liability for AI Hallucinations: Reconsidering Secondary Infringement Standards for Generative AI Systems

Ashmi Jha (NIDAN IPR Research Series 2026)

1.  Introduction

The rapid deployment of large-scale generative AI systems has created legal problems that existing intellectual property regimes were not designed to address. Prior technological disruptions typically generated friction at identifiable legal boundaries digital reproduction, peer-to-peer distribution, online hosting. Generative AI is different. It produces outputs that are potentially infringing, probabilistically unpredictable, and entirely autonomous, without any volitional act that current copyright law would recognise as an infringement trigger. Courts and legislatures have responded, where they have responded at all, by stretching existing analogies rather than confronting the underlying doctrinal failure.[1]

Among the more significant legal concerns arising from this technology is the phenomenon of AI hallucination. In copyright terms, these are not merely factual errors: they include instances where a generative model reproduces, closely paraphrases, or reconstructs protected expression drawn from its training data. The outputs are not random noise but probabilistic reconstructions shaped by statistical patterns encoded during training.[2] Where those patterns encode copyright-protected material, the resulting output may constitute infringement yet identifying who bears legal responsibility for that output is far from straightforward under existing doctrine.

India’s Copyright Act 1957 and the Information Technology Act 2000 were conceived for environments where human agency was a prerequisite to infringement. Secondary liability frameworks, which attribute responsibility to those who facilitate or profit from another’s primary infringement, similarly presuppose identifiable human conduct knowledge, control, inducement. These requirements sit uneasily alongside the operational reality of generative AI, where infringing outputs emerge from automated inference processes that no individual human directs or anticipates. The intermediary liability regime, itself a product of a different technological era, is anchored in notice-and-takedown logic that cannot accommodate outputs which are often unforeseeable and practically irreversible once generated.

This paper argues that existing copyright and secondary liability frameworks are inadequate, both doctrinally and in practical enforcement terms, to address infringing outputs produced autonomously by generative AI systems. This inadequacy generates compounding uncertainty: about who bears liability, how platform accountability should be structured, and what governance responses are appropriate. The paper examines the legal significance of AI hallucinations, analyses the current secondary liability architecture and its limitations, identifies the central attribution problems posed by generative AI, surveys comparative developments, and evaluates the policy challenges that follow from continued legislative inaction.

2.  Understanding AI Hallucinations and Infringing Outputs

The term ‘hallucination’, as applied to large language models and diffusion-based image generators, describes outputs that depart materially from factual accuracy or reproduce protected expression without authorization.[3] From a copyright standpoint, the legally significant hallucinations are not mere factual confabulations but instances of near-verbatim or substantially similar reproduction of text, code, or imagery sourced from training datasets. Empirical research has confirmed that models trained on large corpora exhibit memorization behaviors, producing output sequences that reproduce training material with considerable fidelity.[4]

The problem is compounded by unpredictability. Unlike a human author who deliberately copies, a generative model reproduces protected content as a statistical byproduct of its training. The user who prompts the system may be unaware that a protected work is being reproduced; the developer cannot identify in advance which queries will produce infringing outputs; and the content may circulate widely before detection.[5] This gap between the scale of potential reproduction and the practical limits of detection is precisely what existing copyright enforcement mechanisms fail to address.

The memorization problem is not incidental. Models trained to optimize predictive accuracy on human-generated text will, by design, encode patterns from that text with sufficient granularity to reproduce segments of it under appropriate prompting conditions.[6] The copyright question is therefore not whether such reproduction occurs empirical literature confirms that it does but whether the available legal frameworks can attribute and remedy it effectively.

3.  Existing Copyright and Secondary Liability Frameworks

The Copyright Act 1957 grants authors exclusive rights over reproduction, communication, and adaptation of their works.[7] Section 51 defines infringement to include both direct reproduction and the authorisation of such reproduction by third parties.[8] The threshold question is whether AI-generated output constitutes reproduction in the statutory sense requires an analysis that the Act does not provide. No Indian court has addressed whether the concept of reproduction extends to outputs produced through probabilistic inference over statistical representations of protected works.

The originality standard presents a related difficulty. In Eastern Book Company v DB Modak,[9] the Supreme Court held that copyright requires the exercise of skill, judgment, and creative expression. AI-generated outputs, however sophisticated, do not originate from any such exercise. They emerge from inference over learned distributions. This produces a doctrinal asymmetry: AI outputs may not attract copyright protection in their own right, while simultaneously reproducing the expression of the works they were trained on. The law may deny protection to the output while struggling to attribute liability for the underlying reproduction.

US secondary liability doctrine offers two established routes: contributory infringement and vicarious liability. Contributory infringement requires knowledge of infringing activity and a material contribution to it.[10] Vicarious liability requires the right and ability to supervise the infringing activity and a direct financial interest in it.[11] Both standards were developed in the context of hosting and file-sharing platforms, where the infringing acts were committed by third-party users. They do not translate naturally to generative AI, where the infringing output is produced by the developer’s own system.

India’s intermediary liability framework under section 79 of the IT Act 2000 extends conditional immunity to intermediaries that observe prescribed due diligence standards under the 2021 Intermediary Guidelines.[12][13] In Shreya Singhal v Union of India,[14] the Supreme Court held that this immunity depends on actual knowledge of unlawful content, typically conveyed through court or government direction. The problem is that this regime was designed for platforms that host content generated by others, not for systems that generate content themselves. A generative AI developer is not a passive conduit. The infringing output is the product of its own model. Section 79 immunity is therefore analytically inapplicable to these actors, yet no statutory substitute exists.

Safe harbour provisions under the DMCA and the EU E-Commerce Directive share the same reactive premise.[15][16] Both assume that infringing content pre-exists the platform’s involvement and that liability can be avoided by removing it on notice. Where the platform’s own system generates the content, there is no third-party upload to remove, and the reproductive capacity is embedded in the model’s parameters rather than in a discrete, locatable file. These frameworks were not designed for this scenario, and cannot be reinterpreted to cover it without significant extension.

4.  The Central Liability Problem in Generative AI

The attribution of liability for generative AI outputs is not merely a doctrinal difficulty and it is a practical failure of the existing legal framework. Three categories of actor are potentially implicated: developers who train and deploy the model; users who formulate prompts; and, in some deployment architectures, platform operators who provide API or consumer-facing access. Current doctrine attempts to reach these actors through standards calibrated for human agency. None of them fit cleanly.

The case for developer liability is analytically the most compelling, though doctrinally the least developed. A developer who trains a model on copyright-protected material without authorization engages in large-scale reproduction during training.[17] If that training produces a model capable of reproducing protected expression in its outputs, the developer’s role in creating that reproductive capacity is causally proximate to any resulting infringement. The contributory infringement standard requires knowledge and material contribution. A developer who is aware as the research literature confirms they should be that training on protected corpora produces memorization effects, and who nonetheless deploys such a model commercially, satisfies at least the constructive knowledge threshold.[18]

The difficulty lies in the structure of secondary liability itself. In peer-to-peer cases, the platform’s liability was parasitic on identifiable infringing acts by individual users.[19] Where the generative model itself produces the infringing output, the conventional parasitic structure is unavailable. The model cannot be a legal person. The user may not know that the output reproduces protected material. The developer’s relevant act training is temporally remote from the infringing output and may not satisfy the volitional act requirement that direct infringement doctrine demands.[20]

Vicarious liability is no easier to apply. The right-and-ability-to-supervise prong requires ongoing control over the infringing activity.[21] Once a model is trained and deployed, the developer exercises limited real-time oversight of individual output events. The inference process is automated; no human reviews each query-output pair for copyright compliance before it is delivered. This creates a gap: the actor with the deepest systemic responsibility for the model’s reproductive capacity is precisely the actor least able to exercise contemporaneous supervisory control in the way that vicarious liability requires.

The user-as-sole-infringer theory fares no better. Most users have no knowledge of the model’s training data and no practical means of identifying whether a given output reproduces protected expression. Attributing liability exclusively to users would effectively insulate developers who designed and profited from the model’s reproductive capacity, while penalising the party least responsible for creating it.[22]

Scale compounds all of these difficulties. Generative AI systems are trained on datasets comprising billions of documents, reproducing protected works at volumes that vastly exceed any prior copyright dispute.[23] Enforcement mechanisms built around discrete events like litigation, takedown notices, licensing negotiations are simply not calibrated for reproductive capacity embedded in commercial AI infrastructure at this scale. The transaction costs of enforcement fall disproportionately on individual creators who lack the resources to pursue well-funded AI developers through protracted litigation.

The absence of intent presents a further complication. Secondary liability in India and comparative jurisdictions requires, at minimum, constructive knowledge of infringing activity.[24] Whether a developer who knows that its model exhibits memorisation behaviors can plausibly claim ignorance of specific infringing outputs is contested. The active inducement standard articulated in Grokster may offer a more workable route: a developer who markets a model’s reproductive capabilities, or who fails to implement available technical safeguards against memorized outputs, could satisfy an inducement standard without requiring proof of specific intent. Whether Indian courts would be willing to adopt an analogous approach without legislative direction remains uncertain.[25]

5.  Comparative International Approaches

Ongoing litigation in the United States and United Kingdom offers some indication of how existing doctrine may be extended, though it also highlights the limits of judicial improvisation as a substitute for regulatory reform. The complaint in The New York Times Company v OpenAI, Inc alleges that GPT-4 reproduces verbatim passages from published journalism in response to user queries like a practical demonstration that memorization-based reproduction is not a theoretical concern but a documented commercial phenomenon.[26] The case’s significance lies partly in its framing: the plaintiff attributes liability to the developer, not to individual users, asserting that the training process itself constitutes infringement and that the resulting model embeds a standing reproductive capacity. If that framing is judicially accepted, it will represent a significant development in how secondary liability doctrine applies to AI systems.

In Getty Images (US), Inc v Stability AI Ltd,[27] the English High Court was asked to consider whether images generated by Stability AI, bearing visual similarities to copyright-protected photographs, constituted infringement under English law. The case raises a distinct evidentiary problem: AI-generated visual outputs may reproduce the aesthetic or compositional characteristics of protected works without literal pixel-for-pixel copying. Existing substantial similarity tests were not designed to assess probabilistic stylistic resemblance, and it is unclear whether they can be adapted to do so without producing unprincipled outcomes.

The Authors Guild v Google litigation[28] established that large-scale corpus digitization could qualify as fair use where the use was genuinely transformative and did not substitute for the original in the relevant market.[29] That holding is often invoked in defense of AI training, but the analogy is weaker than it appears. Authors Guild involved content that was stored and indexed for search not reproduced in outputs directed at end users. Where a generative model produces outputs that functionally substitute for the protected work, the market substitution analysis points in the opposite direction, and the transformative use defence becomes considerably harder to maintain.

At the regulatory level, the EU AI Act imposes transparency and training data disclosure obligations on providers of general-purpose AI models.[30] These obligations do not directly create copyright liability, but they represent the first binding acknowledgement in positive law that training data composition is a matter of legal and public concern. The UK Intellectual Property Office, in its 2023 consultation response, retreated from an earlier proposal to expand the text and data mining exception to cover commercial AI training, opting instead for voluntary licensing frameworks.[31] This retreat reflects a wider pattern of regulatory caution one that tends to benefit developers who are already in the market and disadvantages rights holders who lack the commercial leverage to negotiate meaningful licensing terms.[32]

6.  Regulatory and Governance Concerns

The absence of clear regulation in this area reflects both the speed of technological deployment and the institutional difficulty of legislating for a technology that was barely commercially available when most relevant statutes were last substantially revised. India’s Copyright Act 1957 and IT Act 2000 predate not only generative AI but the commercial internet era that shaped them. The resulting regulatory gap is not simply a matter of legislative lag; it reflects a categorical mismatch between the assumptions embedded in existing frameworks and the operational realities of AI-generated content.

Industry actors frequently invoke the innovation argument to resist copyright constraints on AI training and deployment. The concern is not trivial. Premature or poorly calibrated regulation could inhibit AI development in ways that carry real social costs, including in research and accessibility applications. The argument, however, does not justify exempting AI developers from copyright obligations that every other commercial actor is expected to observe.[33] The real regulatory challenge is distinguishing between interventions that impose genuinely disproportionate costs and those that simply require developers to account for the harm they impose on rights holders by training on protected material without compensation.

India’s own legislative process has been slow. The Parliamentary Standing Committee on Commerce’s 2021 review acknowledged the need for policy attention to AI-generated works but did not recommend specific amendments.[34] This caution is understandable in a domain that is technically complex and rapidly evolving, but it has a practical consequence: rights holders are currently left to litigate under frameworks that were not designed for their claims, against well-resourced defendants, at enforcement costs that most individual creators cannot sustain.

The enforcement problem is particularly pronounced in the Indian context. The IT Act’s intermediary liability regime, as interpreted in Shreya Singhal,[35] conditions platform obligations on actual notice — typically a court order or governmental direction. This reactive model cannot keep pace with generative AI: by the time a rights holder identifies and acts on a specific infringing output, the same model may have produced thousands of similar outputs. A notice-and-takedown system designed for discrete content uploads cannot function as a meaningful regulatory response to reproductive capacity embedded in model parameters.[36]

Regulatory divergence across jurisdictions creates additional complications. If some countries impose copyright obligations on AI developers while others do not, the result is a predictable pattern of forum selection and compliance arbitrage that may undermine any single jurisdiction’s regulatory effort.[37] This points toward the need for coordinated responses whether through WIPO’s ongoing AI and intellectual property consultations or through broader treaty-level harmonization of training data disclosure and licensing requirements. Without such coordination, national measures may impose costs primarily on domestically incorporated entities while leaving foreign-based developers effectively insulated.

A separate concern is the opacity of AI systems. Rights holders generally cannot determine whether their works were included in a training dataset without voluntary disclosure from the developer, and cannot assess the extent of memorization without access to model weights or training logs.[38] The EU AI Act’s disclosure requirements are a step in the right direction, but disclosure alone is insufficient unless accompanied by enforcement mechanisms that allow rights holders to act on what is disclosed. Requirements that generate information without enabling its use in legal proceedings are unlikely to provide meaningful protection.

7.  Conclusion

The analysis in this paper leads to a conclusion that is, at once, fairly clear and politically difficult: existing copyright doctrines are not merely poorly suited to generative AI — they are incapable of addressing, without significant extension, the attribution, enforcement, and regulatory challenges it creates. Secondary liability frameworks built on human agency, volitional reproduction, and reactive notice mechanisms cannot be stretched to cover autonomous generative systems without producing legal reasoning that undermines the broader coherence of the liability framework.

The question is not whether revised liability standards are needed. It is whether legislatures and courts will act before the doctrinal vacuum is filled by litigation outcomes that do not reflect considered policy judgments. The cases currently before US and UK courts demonstrate that existing doctrine is being applied to circumstances it was never designed for. Judicial improvisation in this space carries real risks both of over-extending liability in ways that constrain legitimate AI development, and of under-extending it in ways that effectively permit systematic copyright infringement on a commercial scale.

For India, the urgency is compounded by the simultaneous inadequacy of two central statutes. Neither the Copyright Act 1957 nor the IT Act 2000 provides a workable framework for attributing liability for AI-generated infringing outputs, and the intermediary liability regime excludes from its scope the very actors that is the AI developers whose design decisions give rise to the harm. Legislative reform is warranted. Any such reform should draw on the EU AI Act’s approach to training data transparency and consider calibrated secondary liability standards along the lines of the Grokster inducement doctrine standards oriented toward systemic reproductive capacity rather than discrete infringing events.

The most consequential decisions in the generative AI copyright landscape what to train on, what safeguards to build, what reproductive capacities to embed in deployed systems are made by developers, not by users, intermediaries, or regulators. A liability framework that cannot reach those decisions is one that tolerates, by design, the very harm it claims to address.

[1] The New York Times Co. v. Microsoft Corp., No. 1:23-cv-11195 (S.D.N.Y. filed Dec. 27, 2023).

[2] Getty Images (US), Inc. v. Stability AI Ltd., [2023] EWHC (Ch) 3090 (Eng.).

[3] Gary Marcus & Ernest Davis, Rebooting AI: Building Artificial Intelligence We Can Trust 8 (Pantheon Books 2019); see also Ziwei Ji et al., Survey of Hallucination in Natural Language Generation, 55 ACM Computing Surveys 1 (2023).

[4] Nicholas Carlini et al., Extracting Training Data from Large Language Models, in 30th USENIX Security Symposium 2633 (2021).

[5] Andres Guadamuz, Do Androids Dream of Electric Copyright? Comparative Analysis of Originality in Artificial Intelligence Generated Works, 2017(2) Intellectual Property Quarterly 169, 174 (2017).

[6] Complaint ¶¶ 103–119, The New York Times Co. v. Microsoft Corp., No. 1:23-cv-11195 (S.D.N.Y. Dec. 27, 2023).

[7] Copyright Act, No. 14 of 1957, § 14, INDIA CODE (1957).

[8] Id. § 51(a)(i)–(ii).

[9] Eastern Book Co. v. D.B. Modak, (2008) 1 SCC 1, ¶ 14 (India).

[10] Sony Corp. of Am. v. Universal City Studios, Inc., 464 U.S. 417, 435 (1984).

[11] Fonovisa, Inc. v. Cherry Auction, Inc., 76 F.3d 259, 262 (9th Cir. 1996).

[12] Information Technology Act, No. 21 of 2000, § 79(1), INDIA CODE (2000).

[13] Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021, r. 3 (India).

[14] Shreya Singhal v. Union of India, (2015) 5 SCC 1, ¶¶ 122–124 (India).

[15] 17 U.S.C. § 512(c).

[16] Viacom Int'l, Inc. v. YouTube, Inc., 676 F.3d 19, 30–31 (2d Cir. 2012).

[17] Copyright Act, No. 14 of 1957, § 51(a)(ii), INDIA CODE (1957); see also Raman Mittal, Generative AI and Copyright Liability in India, 16(1) Journal of Intellectual Property Rights 44, 47 (2024).

[18] Mark A. Lemley & Bryan Casey, Fair Learning, 99 Texas Law Review 743, 762 (2021).

[19] Fonovisa, 76 F.3d at 263; Metro-Goldwyn-Mayer Studios Inc. v. Grokster, Ltd., 545 U.S. 913, 937 (2005).

[20] Pamela Samuelson, Generative AI Meets Copyright, 381 Science 158, 159 (2023).

[21] Fonovisa, 76 F.3d at 263.

[22] Samuelson, supra note 20, at 159.

[23] Carlini et al., supra note 4, at 2640.

[24] Grokster, 545 U.S. at 940.

[25] Id.

[26] Complaint ¶¶ 55–63, The New York Times Co. v. Microsoft Corp., No. 1:23-cv-11195 (S.D.N.Y. Dec. 27, 2023).

[27] Getty Images (US), Inc. v. Stability AI Ltd., [2023] EWHC (Ch) 3090, ¶¶ 44–51 (Eng.).

[28] Authors Guild v. Google, Inc., 804 F.3d 202, 212 (2d Cir. 2015).

[29] Id. at 216–17.

[30] Regulation 2024/1689, of the European Parliament and of the Council of 13 June 2024 Laying Down Harmonised Rules on Artificial Intelligence, arts. 50, 53, 2024 O.J. (L 1689) (EU).

[31] U.K. Intellectual Property Office, Artificial Intelligence and Intellectual Property: Copyright and Patents—Government Response to Consultation 14–18 (2023).

[32] Id. at 22.

[33] Ryan Abbott & Elizabeth Rothman, Disrupting Creativity: Copyright Law in the Age of Generative Artificial Intelligence, 75 Florida Law Review 1141 (2023).

[34] Department-Related Parliamentary Standing Committee on Commerce, Rajya Sabha, Review of the Intellectual Property Rights Regime in India, 161st Report ¶ 6.14 (2021).

[35] Shreya Singhal v. Union of India, (2015) 5 SCC 1, ¶ 119 (India).

[36] Information Technology Act, No. 21 of 2000, § 79(3)(b), INDIA CODE (2000).

[37] Lilian Edwards & Michael Veale, Enslaving the Algorithm: From a “Right to an Explanation” to a “Right to Better Decisions”?, 16(3) IEEE Security & Privacy 46, 48 (2018).

[38] Samuelson, supra note 20, at 161.

Suggested Citation

Ashmi Jha, “Copyright Liability for AI Hallucinations: Reconsidering Secondary Infringement Standards for Generative AI Systems,” NIDAN Centre for Legal and Policy Research (2026).