In light of the first few issues framed by the Court in ANI v. OpenAI copyright dispute, Shama Mahajan in this two part post discusses the problem of excessive judicial borrowing and fair use defence vis a vis storage. In Part I of the post, she explains non-applicability of ‘derivative work’ theory in the Indian context owing to the lack of its statutory sanction. In Part II of the post, she argues that the Fair Use defence of incidental or transient storage might not be the right way out given the dynamics of how the data processing and storage work in these models. Shama is an LL.M Candidate at National University of Singapore, pursuing her masters in Intellectual Property and Technology Law.
ANI v. Open AI – A Lesson in Resisting the Temptation to Borrow Excessively without Legislative Sanction
By Shama Mahajan
The Delhi High Court is currently presiding over one of the most anticipated disputes of the AI-age, ANI v. Open AI whereby the traditional realms of copyright law are at loggerheads with the dynamic functionality of the AI machine learning systems. The dispute has been covered on the blog here.
The Court has framed the following issues on the interim injunction application by ANI:
- Whether the storage by the defendants, of plaintiff’s data (which is in the nature of news and is claimed to be protected under the Copyright Act, 1957) for training its software i.e., ChatGPT, would amount to infringement of plaintiff’s copyright.
- Whether the use by the defendants of plaintiff’s copyrighted data in order to generate responses for its users, would amount to infringement of the plaintiff’s copyright.
- Whether the defendants’ use of plaintiff’s copyrighted data qualifies as ‘fair use’ in terms of Section 52 of the Copyright Act, 1957.
- Whether the Courts in India have jurisdiction to entertain the present lawsuit considering that the servers of the defendants are located in the United States of America.
For the purpose of this post, I will focus only on the first 2 issues with the objective of highlighting this instance as an opportunity for the Court to stir clear from repeating the mistake of ‘excessive borrowing’ of judicial principles from other jurisdictions in absence of corresponding legal provisions. A lot has already been written and discussed as far as the intricacies of the technology and law involved in this issue are concerned. This post focuses on the repetitive mistake that Indian courts have been committing since the judgement of EBC v. D.B. Modak and even in this preliminary stage, the Delhi High Court seems to have already started its journey on that path.
The post highlights why such approach is detrimental and the underlying flaw in its application to the present case. It aims to highlight that, borrowing must be cautionary in this particular case and if so done, can set in motion the new jurisprudence with this verdict.
EBC v. DB Modak: The precedent that went beyond Originality
The landmark decision of the Supreme Court of India, in EBC v. DB Modak laid down the standards for determining ‘originality’ for a work to be granted copyright protection. However, it also went beyond to establish the Indian standards for ‘derivative work’ and the threshold of when would a ‘derivative work’ be considered infringement and when would it be entitled for a separate copyright protection. The Courts created the primary/prior works and secondary works classification without analysing whether Indian law envisages, requires and justifies such artificial classification. It is to be noted that, the Apex Court was heavily swayed by the jurisprudence of US Courts without paying due heed to the fact that judicial precedents are developed on the background of a statutory provision.
Derivative Work: Statutory Creation v/s Judicial Creativity
In the US, § 106 of the Copyright Act explicitly recognises ‘derivative work’ as an exclusive right of the copyright owner. Thus, the famous test of ‘transformativeness’, was a judicial creation for determining the statutory right in the context of fair use.
There exists no such corresponding provision under the Indian Copyright Act, 1957 (Indian Copyright Act), which seeks to confer any right to create derivative work upon the copyright owner. Thus, the artificial classification of prior work and secondary work with an implied test of transformativeness to determine the copyrightability and fair use, does not stand the test of Indian Copyright Act.
Derivative Work: Isn’t it an Adaptation?
An argument may be made, that derivative work under US law is reflected in the Indian context as the ‘right of adaptation’. However, this argument does not hold true for the following reasons:
- If the intent of the Supreme Court was to equate derivative work with an adaptation, the Court could have easily drawn the parallel which is absent in the DB Modak Case
- Derivate Work under the US Law is an all-encompassing right without any restriction on the type of underlying work i.e. it is a right not determined by the subject-matter of copyright. Under Indian Copyright Act, adaptation is an exclusive right which is available only to a specific category/subject-matter of copyright i.e. literary, dramatic, musical and artistic work (Section 14).
- The scope of adaptation under the Indian Copyright Act is well-defined u/s 2(a) whereas derivative work in the US context is intended to be broader with the objective of excluding any and all means which expose the primary work even if in fragments to the public.
Thus, the intent of the Indian Supreme Court was to create an overarching principle of ‘derivative work’ transcending the scope of adaptations.
The ‘Derivative Work’ Paradox: A Venus Fly Trap
The second issue framed by the Delhi High Court is an inevitable analysis of ‘Derivate Work’ given that, there is hardly any distinction between ‘storing copyrighted material for training’ vis-à-vis ‘using copyrighted material to generate response’. Even in the statutory sense, the ambit of reproduction includes storage as per section 14(a)(i) of the Indian Copyright Act. It is thus evident that, the Court intends to examine whether output created by Gen-AI models like ChatGPT can be qualified as derivative work without infringing the copyright of the original copyright owner for being sufficiently transformative, and thus protected by fair use.
It is to be noted that, the jurisprudence of ‘transformativeness’ is very strongly litigated and hence established (though highly debated) in the US context with cases like Cariou v. Prince and Andy Warhol. It is also important to note that the US Court’s discussion on transformativeness to qualify as fair use has evolved in its approach. Initially, it was centered around ‘whether the defendant intended to alter the copyrighted work’s expression, meaning or message’. In recent times the court of appeals for the second circuit, inciting disagreement from the seventh circuit, has held that, the intention of the defendant itself is not sufficient to justify transformativeness, for what matters is how the work appears to a reasonable observer. It is not only this contrast but also the growing criticism of the evaluation of transformativeness as an uncertain standard resulting in at times justifying even exact copies as fair use which is highlighted by Prof. Rebecca Tushnet in her paper, that raises the question of whether the Indian jurisprudence on the same is equally developed to handle the Gen-AI angle.
Indian jurisprudence on transformativeness, mostly reflects fragments of the US understanding tailored to fit Indian scenarios. In the cases where transformativeness has been relied upon to uphold fair use, the analysis was limited to ‘transformation of the purpose by the subsequent work than the original’. Thus, Indian analysis is more content-driven focusing on ‘objective replication or copying of the form of expression of the original work without transformative purpose’, whereas US Courts examine intent of the creator of the secondary work along with the purpose hence, it is more intent-driven as observed by various scholars including Tushnet and Anthony Reese. The complication can be best highlighted by the dichotomy of how a case like Warner Brothers Entrainment Inc. v. RDR Books involving Harry Potter Encyclopedia would be treated in India in the light of the Indian precedent of University of Cambridge v. B.D. Bhandari when in the US it was held to be non-transformative. As per the Indian standard there is no objective replication or copying of the original expression and the purpose served by the encyclopedia is substantially different.
Another issue that also arises in application of ‘transformativeness’ to assess fair use in Indian context is that, since right of adaptation is more limited than the right to create derivative work, transformativeness in its letter and spirit would then run the risk of encompassing rather than carving out an exception under fair use. In other words, permitting exceptions that are wider than the right itself which would go foul of the statutory object and purpose.
The Court if it chooses to enter this discussion will open a series of permutations and combinations whereby instance-based determination of transformativeness will be required which will become difficult as predicting output of an AI system to a prompt is difficult. This approach in my opinion is misleading in the Gen-AI context of the Indian statutory landscape and will stir the discussion on an entirely different tangent.
The Statutory as The Primary
In my opinion, the absence of ‘derivative work’ as a statutory provision must be respected and the Indian Courts must limit themselves within the statutory provisions while analysing the same. The existing Indian Copyright Act, and its provisions are sufficient to determine the questions of law involved in this matter. In my opinion, the material issues (barring jurisdiction) should have been framed in the following manner:
- Whether storage and use by the defendants of plaintiff’s data for training its software i.e., ChatGPT would constitute reproduction and thus amount to infringement of plaintiff’s copyright.
- Whether output generated by ChatGPT would constitute either reproduction and/or making of copies and/or adaptation of the original work thereby infringing the copyright of the plaintiff.
- Whether the defendants’ use of plaintiff’s copyrighted data for Issue No. 1 and/or Issue No. 2 qualifies as ‘fair use’ in terms of Section 52 of the Indian Copyright Act.
In the above issues, the Court would be in a position to acknowledge the Gen-AI supply chain catering to the crucial phase distinction of the Gen-AI Models i.e. training (pre-training + fine tuning) and deployment or generation (as shown below).
Focusing on the said issues will allow the Court to discuss more intricate questions relating to storage of data, the manner in which it is stored, the possibility of training without storing the data on centralized cloud servers etc.
Efficient Borrowing: An aid to solving, not the solution!
In absence of ‘derivative work’ analysis, the Court will be limiting itself to the statutory provisions. I also feel that, without ‘derivative work analysis’, examining the underlying questions of using and storing data for pre-training and subsequent analysis of the output will center around ‘idea and expression’ related discussions which are more central to this issue given the manner in which Gen-AI systems use the content it is trained on.
In the aforementioned issues, the court can evaluate issues like transient storage of copyrighted expressions during training of Gen-AI models and whether the same can be considered an infringement. It is not my intention to argue that borrowing must be avoided, rather to say that borrowing must be secondary and cautionary. For example, in analysing transient storage of data if and when it happens during the training period, the Courts will benefit to draw from the precedents like Cartoon Networks v. CSC Holdings in addition to the Section 52(1)(b) and (c) of the Indian Copyright Act along with the Indian precedents like MySpace Inc. v. Super Cassettes Industries Ltd. Therefore, learnings from another jurisdiction must first be tested on the anvil of the statutory provisions to ensure seamless integration.
In Part II of the post, I’ll discuss the first issue of storage and how fair use defence for incidental and transient storage might not be the right way to go about.