An AI Tax in the Name of Copyright?  Analysis of Remuneration Proposals in India

Background

A committee appointed by the Department for Promotion of Industry and Internal Trade (DPIIT), which oversees the Copyright Office in India, has released the first part of a working paper on Generative Artificial Intelligence (“GenAI”) and Copyright. The paper titled ‘ONE NATION, ONE LICENSE, ONE PAYMENT – Balancing AI Innovation and Copyright’ (“DPIIT paper”) focuses on the copyright issues relating to AI training and advocates the introduction of a new licensing framework to facilitate AI training in India. The second part of the paper, which will be released later, is expected to address authorship and copyrightability related questions of outputs generated by GenAI. So far the Indian government was largely refraining from taking any positions on copyright issues related to AI training. But as the paper is coming from a government appointed committee, it is expected to have substantial impact on future policy making in the area. In view of the fact that India is already witnessing copyright-infringement-related litigations relating to AI training, the paper could even have an impact on the ongoing and future litigation in the area.  

Core aspects of the proposed model

The DPIIT paper proposes a hybrid model of a mandatory blanket license, complemented with a statutory remuneration right. The proposed model allows the use of all lawfully accessed copyright-protected works for AI training and the copyright holders will not have an opt-out option. The model envisages a centralised entity named Copyright Remuneration Collective for AI Training, composed of copyright societies and collective management organisations (CMOs), to collect payments from the developers of the AI systems and distribute the money among rightsholders, including non-members of those organisations who have registered their works for receiving royalties for AI training. 

According to the paper, the core philosophy behind the suggested model is “to address the challenges of large-scale data use by AI Developers while ensuring that creators are fairly compensated for the use of their works.” They also add that the model would “ensure automatic availability of copyright-protected works for training of AI Systems and legal certainty, to help unlock the transforming potential of AI Systems for mankind.”

Some observations regarding the proposed model

On the positive side, the paper paves way for more concrete policy discussions in India on the critical question of allowing the use of copyright protected materials for AI training. However, the proposed model also raises several concerns. This post focuses on three such areas that may not have received due attention.

Methodology 

A major limitation of the DPIIT paper, and the hybrid model it recommends, is the methodology adopted to arrive at the recommendations. As is evident from different parts of the paper, it takes a lawyer’s approach, rather than a researcher’s approach, for addressing core issues. In other words, the discussion in the paper moves with many pre-drawn conclusions and engages in backward reasoning to justify the conclusions. This is primarily done through over-emphasis on literature and case-laws that support certain view points, while ignoring case-laws and literature that goes against the views.

For instance, the authors of the paper state at page 22 that 

“[t]his working paper, therefore, does not attempt to resolve these questions or offer definitive conclusions on whether infringement is made out and/or the “fair dealing” exception applies. Instead, the goal is to propose a forward-looking legal and policy framework.” 

However, a closer examination of the discussion on infringement and exceptions to infringement (pages 13-22), suggests that the authors are in effect trying to convey that there is very high probability of finding infringement in the context of AI training and that the scope of exceptions under Indian copyright law is very narrow. This view is made more explicit with the following statement: 

“As discussed above, there are strong reasons to believe that training AI Systems may raise concerns about copyright infringement. Simultaneously, there is an ongoing debate regarding whether the “fair dealing” exception can be applied to the process of training of AI Systems” (page 22). 

The DPIIT paper reaches this conclusion primarily by avoiding engagement with literature that points out that AI training may be permissible under the current version of Indian copyright law, if one looks at various case-laws on infringement and exceptions to infringement. At places, superficial comparative analyses also contribute to that objective. For example, in the context of the discussion on fair dealing, the paper refers to Canada and notes that: “Canada also has a similar, narrower “fair dealing” exception, and the Canadian Supreme Court has rejected the applicability of the broader American “fair use” test.” While it is true that Canada has rejected the applicability of the American fair use test, what the paper conveniently ignores is the broad manner in which the Canadian Supreme Court has interpreted fair dealing provision paving the way for a much stronger user rights jurisprudence under copyright law.

Yet another area wherein the methodology-related issues are evident is the presumption of market harm. As discussed by Akshat Agarwal, no part of the paper provides any empirical evidence from any sector within or outside India on potential underproduction of works due to GenAI. It also fails to engage with literature that suggests how any such market harm (if existing) could be addressed. It is worth highlighting here that in litigations in other parts of the world, particularly the decision in Kadrey vs Meta Platforms, Inc., one can see the plaintiffs failing to produce any empirical data despite many strong nudges from the side of the court. This reaffirms that it is highly problematic to assume market harm and it would be a terribly bad step to suggest major policy steps based on weak assumptions. 

While the approach taken in the paper might be permissible, and may be even desirable, to protect client interests within a court room, a working paper from an official committee of the government should ideally be avoiding  a lawyer’s approach to drawing conclusions and recommendations. As is evident from a compilation of footnotes in the DPIIT paper, it is also disappointing to see that almost all of the 43 academic works cited in the paper are from authors located outside India. This should certainly be an epistemological concern for a country that wishes to set the agenda for the Global South in AI related dialogues!

Lawful access

One of the other areas wherein the proposed model may cause substantial harm in the long run beyond the realm of AI training is the emphasis on “lawful access” as a pre-requisite for the use of copyrighted works for training (Page 62). 

What is clearly missing in the paper is clarity on what “lawful access” means in the context of AI training. A preliminary reading of the paper gives the impression that the intention of the authors is to limit it to avoiding circumvention of technological protection measures to access copyrighted contents. But the paper ignores certain core ambiguities on the idea of lawful access. For instance, if someone has personal subscription to a database of law review articles, wherein the terms and conditions of access specifically allows the user only to read and download articles for personal purposes, an important question is whether the use of materials from that database for AI training can be considered as “lawful access”. This issue would be even more complicated if the database provider also provides a separate license (with a substantially high fee structure) for AI training purposes. Similar questions could also be posed with respect to other types of works such as musical works. 

It is also important to note that if one looks at the current exceptions framework under Indian copyright law, it is evident that the lawful access requirement is not a general requirement to exercise the user rights under Sec. 52 (the exceptions provision under Copyright Act 1957) and that it has been historically limited to certain very specific use scenarios. For example, Sec. 52(1)(aa), Sec. 52(1)(ab), Sec. 52(1)(ad) and the explanation to Sec. 52(1)(a) of Copyright Act 1957 indicate the intention of the legislators to have some variants of the lawful access requirement in specific use scenarios relating to computer programmes. But the legislators have not suggested a similar requirement for other kinds of works or most use scenarios covered under Sec. 52. This is not without a reason, as such an obligation would have unduly curtailed the scope of user rights under Indian copyright law and would bring in substantial uncertainties for all stakeholders.  In this context it is also worth adding here that even in a country like the US, the position is not well settled with regard to the question of lawful access to claim fair use.

The broader point here is that the idea of ‘lawful access’ is highly subjective and the DPIIT paper overlooks the fact that it is a slippery slope when it comes to the exercise of user rights under copyright law. By sneaking in the lawful access requirement in the context of AI training without substantially engaging with the complexities and the potential negative consequences of such a pre-requisite, the paper is inadvertently promoting more ambiguities and uncertainties in the area of copyright law. This is even more worrisome in the context of the model suggested in the paper, as it also mentions that the primary burden of proof with regard to compliance is on AI developers (page 61).

Can a license be imposed for potentially non-infringing uses?

In almost all other cases wherein statutory licensing or compulsory licensing have been provided under copyright law, it is evident that in the absence of taking such a license, there will be potential infringements of the rights of copyright holders. For example, when a radio station plays a song without taking permission from the copyright holders, it is evident that there will be infringement of copyrights. To address that issue with minimal transaction costs for all stakeholders and to ensure fair compensation to copyright holders, a statutory license or compulsory license is envisaged by law makers in many jurisdictions. Sec. 31D of the Copyright Act 1957 in India, which provides statutory licence for broadcasting of literary and musical works and sound recording, is an example in this regard. But in the case of use of copyrighted works for AI training purposes, it is not yet clear in most jurisdictions, particularly India, whether there is any copyright infringement at all. As indicated above, the DPIIT paper has avoided engaging meaningfully with this fundamental question. If there is no violation of copyright when works are used for the purpose of training, then the imposition of a statutory licensing model is going to tilt the balance unreasonably in favour of copyright holders, without any legal or economic basis. This is not to say that the government cannot impose a tax or cess for such uses. But framing it as something akin to a statutory license might be a legally weak proposition. 

Other issues

As indicated above, there are also other major conceptual challenges within the paper, as pointed out by many other scholars like Kalyan C Kankanala, Swaraj Barooah and Rahul Matthan. For example, in the name of promoting AI innovations, the suggested model is in effect also taking away the autonomy of the authors to prevent their works from being used for training purposes. The fact that this has been done without adequate legal or economic reasoning makes it highly problematic. It is also important to ask whom the proposed model will benefit. Though the authors of the paper have carefully avoided providing any estimates of potential revenues and distribution among different stakeholders, a close perusal of the distribution model suggested will convey that it is only the large copyright holders (like big music labels and major publishers) who will get any economic benefit from the proposed model. Individual artists and authors are not going to have any bargaining power or meaningful economic remuneration under the proposed system.

Conclusion

The authors of the DPIIT paper deserve appreciation for producing a comprehensive analysis within a very short period of time, and thereby also paving way for initiating more conversations with the Indian government on copyright related issues relating to AI training. However, though the paper is well-intentioned, there are several areas of concern in the proposed hybrid model. The model, if implemented, can best be characterised as a tax on AI developers, rather than a copyright licensing model. While the redistribution of money through taxes on big tech companies has been proposed by many scholars to address the increasing challenge of perpetuation of inequalities, and to support universal basic income, the fundamental question here is whether we should do it in the name of copyright! Maybe we should just call it an AI tax and find a broader and equitable wealth redistribution mechanism that benefits all sections of the society, and not just a few big copyright holders!

{Prof. Arul George Scaria is a professor of law at the National Law School of India University, Bengaluru and co-director of the Centre for IP Research and Advocacy}