Beyond Liability: Rethinking Copyright Remedies in ANI v. Open AI

[Addepalli Aaditya and Srijan are students at Hidayatullah National Law University.]

ANI Media Private Limited v. OpenAI OpCo LLC (ANI v. OpenAI) at the Delhi High Court (Court) asks whether using ANI’s online news articles to train ChatGPT constitutes copyright infringement under Indian law. Most of the debate so far has focused on liability, for instance, whether large-scale scraping and storage of news content is “reproduction” under the Copyright Act 1957 and whether OpenAI can rely on “fair dealing” exceptions in Section 52.

ANI has also demanded that OpenAI be prohibited from accessing the platform as well as using the contents for training its models, and that ChatGPT be barred from accessing the ANI website while it is seeking remedies. The Court, after over a year of arguments over ANI’s interim application, has reserved its judgment, but has yet to determine what, if anything, should be done in response. Given that ANI’s pleadings primarily concern liability, the Court’s present focus is unsurprising. However, the more significant question remains what remedies would be available if infringement is ultimately established. This article examines the post-liability remedial questions that will likely be raised if infringement is established in the future through the lens of ANI v. OpenAI. When the work has been incorporated into a large-scale foundation model like ChatGPT, traditional copyright remedies do not map neatly onto AI training. Indian and foreign courts are already adapting existing doctrines to this AI-training context, but there is no settled remedial framework yet.

The Existing Remedial Architecture Under Indian Copyright Law

Upon establishing infringement under the Copyright Act 1957, the copyright owner may be able to seek common civil remedies. Section 55 provides that the owner “shall…be entitled to all such remedies by way of injunction, damages, accounts and otherwise as are…conferred by law”. In practice, courts may award damages or direct an account of profits depending on the circumstances of the case. The proviso to Section 55 restricts “innocent infringer” relief. Where the defendant proves a lack of knowledge of the copyright, the plaintiff may be confined to injunctive relief and, in appropriate cases, an account of profits rather than full compensatory damages.

Section 58 specifically deals with infringing copies. It enables the copyright owner to seek delivery up, detention or destruction of infringing copies, thereby treating such copies as the property of the copyright owner for remedial purposes. Under Section 58, the copyright owner can essentially treat infringing copies (whether in book form, disc or identifiable digital file) as its property and have it delivered up or destroyed. However, there are only limited protections for the innocent purchaser. Separately, Order XXXIX, Rule 1 and 2 of the Civil Procedure Code 1908 states that the injunction may be issued if the following conditions are satisfied: (a) the plaintiff has a prima facie case, (b) the balance of convenience is in favour of the plaintiff, and (c) there exists a risk of irreparable harm if the injunction is not granted. In Wander Limited v. Antox India Private Limited, the Supreme Court clarified that even where a prima facie case and irreparable injury are established, an injunction may still be refused if the balance of convenience favours the defendant.

The rules, viewed collectively, appear to restate the conventional copyright enforcement model. The framework assumes the existence of an identifiable infringing copy, that a continuing copying, such as sale or distribution, can be prohibited, and that the plaintiff’s loss and/or the defendant’s gain can be quantified in monetary terms. Provided an act is proven, the remedies will be customised. For instance, where pirated DVDs are being manufactured or distributed, courts may seize the infringing copies, restrain further circulation and award damages or an account of profits.

How Foundation Models Strain This Framework

Foundation models like ChatGPT strain three assumptions behind Sections 55 and 58 and Order XXXIX. First, they challenge the idea of an “identifiable copy”: Section 58 appears better suited to tangible or readily identifiable copies, but it remains unclear whether the trained model itself can be characterised as an “infringing copy” within the meaning of the Copyright Act 1957. Before the Court, OpenAI has submitted that ANI’s site was blocklisted in October 2024 and would be excluded from future model training. It also maintained that ANI’s articles are not stored in a human-readable or directly retrievable form within its trained model, and US plaintiffs have even asked courts to destroy allegedly infringing models.

Second, they strain the “reversibility” idea behind injunctions. Section 55 and Order XXXIX are meant to stop ongoing or future wrongs, not undo past training. So, courts can target future access, deployment, or retraining, but not easily “unlearn” a finished model.

Third, they strain proportionality. Destroying or disabling a widely used model could be significantly disruptive, so any remedy should be narrow and tied to actual harm. That is why commentary on ANI v. OpenAI stresses targeted injunctions and calibrated money awards rather than blanket relief.

Comparative Remedial Approaches

Although no jurisdiction has yet developed a comprehensive framework for such AI copyright disputes, a few decisions illustrate the different ways in which traditional remedies are being adapted to foundation models.

The UK High Court, in Getty Images v. Stability AI, dealing with the identifiable-copy problem, held that Stability AI’s model weights were not an “infringing copy” capable of constituting Getty’s secondary infringement claim under the Copyright, Designs and Patents Act 1988. The decision highlights a fundamental limitation of conventional copyright remedies, which ordinarily presuppose an identifiable infringing embodiment. Where the alleged infringement has been absorbed into a training model rather than retained as a discrete copy, remedies such as ‘delivery-up’ become harder to materialise. This raises a similar question for Indian law, which, under Section 58, operates on the premise that infringing copies can be identified and treated as the copyright owner’s property.

Litigations such as The New York Times v. OpenAI, in the US, also remain pending, leaving both liability and remedies unresolved. However, American scholars have begun exploring alternatives such as the license-fee theory of damages, which seeks to measure the value derived from unauthorised training rather than placing sole reliance on traditional market-loss calculations. Likewise, a permanent injunction against AI developers continues to be governed by the equitable test laid down in eBay Inc. v. MercExchange. This test requires courts to balance the necessity and practicality of injunctive relief and not assume that every proven infringement warrants such a relief.

The EU takes a different approach. Rather than addressing the remedial questions post-infringement, Article 4 of the Digital Single Market Directive narrows the scope of infringement itself by permitting text-and-data mining for commercial purposes unless right-holders expressly opt out. In doing so, the need for courts to articulate complex post-liability remedies is reduced.

The learnings from these jurisdictions lie in demonstrating that the core challenge is remedial, not doctrinal. As Indian courts confront similar disputes, the challenge lies in how existing copyright principles can be flexibly applied to technologies for which they were never originally designed.

The Way Forward

Indian courts should not start off by searching for newer AI-specific remedies. Section 55 already offers flexibility through injunctions, damages and accounts of profits, while Order XXXIX requires courts to tailor interim relief to the circumstances of each case. Rather than treating every successful claim of infringement as warranting a broad injunction, courts should assess the form of relief most suited for addressing the harm that has actually been established.

In practice, this means moving away from binary remedial choices. If the continuing use of copyrighted material can be restrained prospectively, narrowly tailored injunctions may be appropriate. Where reversing completed training is not possible, monetary relief may better address the copyright owner’s interests without imposing consequences that extend beyond the infringement proved. At the same time, the application of Section 58 to foundation models should take account of ongoing developments in machine unlearning. Researchers are exploring techniques such as selective forgetting, model editing and retraining on subsets of the training data to reduce or remove the influence of particular works from a trained model while preserving its overall functionality. These techniques remain experimental and are not yet capable of reliably removing individual works from large foundation models.

Nevertheless, continued advances may in time, make it technically feasible for courts to consider remedies requiring AI developers to remove the influence of identified copyrighted works from trained models. If that occurs, Section 58-style delivery-up remedies may become more workable than they are today. Until then, courts should assess the availability of such remedies in light of present technological capabilities rather than speculative future developments.

The remaining problem, however, calls for a more functional understanding of Section 58. Rather than treating delivery-up as a binary remedy, Indian courts should confine its application to identifiable infringing embodiments such as datasets or cached copies. This tries to preserve the underlying principle of Section without forcing the provision to operate where no identifiable copy exists.

ANI v. OpenAI therefore represents more than another copyright dispute involving AI. Regardless of how the Court ultimately decides the question of liability, the case signals that the next phase of AI copyright litigation will turn not simply on whether infringement occurred, but on whether the existing remedial architecture can meaningfully respond once it has.