Source: IDEAS AND INSTITUTIONS | ISSUE #40
Analysis
ChatGPT, Intellectual Property, and Economic Development
In July this year, American stand-up comedian and author Sarah Silverman, along with others, filed copyright infringement suits against OpenAI and Meta. These cases, and others, allege that companies like OpenAI have been using publicly available online material that is copyrighted without requesting any permission, consent, or license from the authors and creators. The outcomes of these cases could very well shape the future of generative artificial intelligence (AI). As this essay speculates, this in turn has implications for how countries like India can benefit from products such as ChatGPT.
As of now, there have not been any definitive judgments in India or elsewhere on the issue of intellectual property and generative AI. However, it may be worthwhile to think through the implications that legal developments will have in shaping this market of generative AI products and the economic consequences for markets like India. In this essay, I pose questions exploring these issues.
What Is Generative AI?
The term generative AI is used for “any artificial intelligence tool that generates something new from existing data when prompts are given, like an image or text.” OpenAI’s ChatGPT, Google’s Bard, and Microsoft’s AI-powered Bing are some of the most prominent examples of generative AI products. Other products include image generators like Stable Diffusion, audio generators such as Resemble AI, and also code generators like Copilot. Indian enterprises have created generative AI products as well; some examples include KissanAI, Bharat GPT, and Jugalbandi.
Generative AI products have become extremely popular. ChatGPT, for example, had over 100 million users in less than three months of its introduction. Since then, OpenAI has also introduced a subscription-based version of ChatGPT. Many generative AI products are similarly available both for free and with additional features for subscribers. This subscription-based model is significant for the ensuing discussion.
But how is generative AI trained? Amazon Web Services’ website explains that “generative AI is powered by very large machine learning models that are pre-trained on vast amounts of data, commonly referred to as foundation models (FMs). A subset of FMs called large language models (LLMs) are trained on trillions of words across many natural-language tasks. These LLMs can understand, learn, and generate text that’s nearly indistinguishable from text produced by humans.”
According to one scholar at Georgetown University, the quality of the product is dependent on three things: the quality and quantity of data, algorithms, and computing power. The data used to train FMs and LLMs is becoming an increasing source of contention.
Generative AI and Intellectual Property Issues
First, a generative AI product uses large amounts of data to be trained. Second, it responds to a user’s input or prompt in a chat box to generate an output. Given this process, there are at least two potential sources of conflict over intellectual property rights (IPR) in a generative AI product.
- Intellectual property issues over the training data.
- Intellectual property issues over the generated output.
The second issue is already being discussed at various forums. The question in this case is: Does someone using a generative AI product to write a story or create art deserve a copyright over his or her creation?
Government agencies, international bodies, and professional associations have already released preliminary statements on these issues. In March this year, the U.S. Copyright Office published a statement to the effect that human authorship is necessary for any grant of copyrights. The World Intellectual Property Organization (WIPO) has also flagged these issues. The International Association for the Protection of Intellectual Property passed a resolution on the same lines as the U.S. Copyright Office, declaring that human intervention is necessary for the grant of a copyright.
While the second issue will continue to be discussed, in this essay, I focus on the implications of the first: are there IPR issues related to the data used to train the FMs that power generative AI products?
Sarah Silverman’s suits claim that there are. Her suits against OpenAI and Meta allege that the datasets on which ChatGPT and Llama were trained used her and her fellow plaintiffs’ copyrighted books without their consent. They allege that “shadow library” websites, such as Library Genesis, Z-Library, and Sci-Hub, were among the datasets used for training the FMs. They further add that “these shadow libraries are also flagrantly illegal.”
The suits also allege that because these generative AI products cannot function without the information extracted from the infringed copyrighted works, the FMs themselves are “infringing derivative works.” In addition, the suits claim that during the process of training, OpenAI and Meta removed the copyright management information, which includes the copyright notice, title, names of owners, and so on. Because of these violations, the suits claim that OpenAI and Meta can also not claim to have any copyright over the language models.
A similar suit was filed against Stability AI, which runs Stable Diffusion. The product generates images in response to text prompts. The suit alleges that to create its training dataset, Stability AI “scraped, and thereby copied over five billion images from websites as the Training Images used as training data. Stability did not seek consent from either the creators of the Training Images or the websites that hosted them.” Further, the suit alleges that “every output image from the system is derived exclusively from the latent images, which are copies of copyrighted images.”
Implications for Generative AI Businesses
There is therefore a set of interesting cases that could lead to extremely consequential outcomes for generative AI products. If we assume for a moment that the suits will be completely successful, that is, a hundred percent of the plaintiffs’ claims in each of these suits are decided in their favor, it will probably lead to the following outcomes and questions:
- Generative AI product companies will have to either seek the required use permissions or licenses or/and pay royalties for every copyrighted or creative work they use for training FMs. How would this be managed? What are the contractual and financial costs this would impose on generative AI businesses?
- In the absence of such permissions or licenses, they would not be able to use copyrighted works for training FMs. Would their datasets then be trained only on works in which no copyright subsists or has already expired?
- Generative AI companies will not be able to remove copyright management information (under U.S. law, at least) from the training dataset or will have to create some mechanism to make sure this information is visible in the output. How will this be implemented in a manner that retains the current “chat box” nature of the products and therefore makes them easy to access?
- On a related note, if generative AI products themselves are considered infringing works because they are trained on infringing data, would this lead to reduced commercial interest in creating the kinds of products available in the market today?
- How would this impact the many companies that are building application-based products using the FMs built by OpenAI, Meta, and Google? If the underlying dataset is held to be copyright infringing, would this lead to significant negative impacts on such businesses?
While it is unlikely that these suits will be completely successful, they may be partially successful. Similar suits down the line may chip away to create more IP rights slowly; some elements of author and creator rights may be recognized gradually in disparate cases or laws. If the training methodology described in the suit is accurate, it will also be hard to argue that companies should be simply permitted to collect pirated books and copyrighted art from the internet and develop commercial products using them, including, as I mentioned earlier, subscription-based products.
At the same time, there is another alternative that should be discussed: Does IPR law need to change because of generative AI and the possibilities it creates? WIPO’s document referenced earlier frames these alternatives well:
(ii) If the use of the data subsisting in copyright works without authorization for machine learning is considered to constitute an infringement of copyright, what would be the impact on the development of AI and on the free flow of data to improve innovation in AI?
(iii) If the use of data subsisting in copyright works without authorization for machine learning is considered to constitute an infringement of copyright, should an explicit exception be made under copyright law or other relevant laws for the use of such data to train AI applications?
…
(vi) Would any policy intervention be necessary to facilitate licensing if the unauthorized use of data subsisting in copyright works for machine learning is considered an infringement of copyright? Would the establishment of mandatory collective management societies facilitate this? Should remedies for infringement be limited to equitable remuneration?
While no clear answers have emerged yet, as is clear, the conflict between the existing ways of developing generative AI products and IPR will be important. Many commentators have, for example, talked about the impact these products will have on employment and the future of work. These take for granted the present design of generative AI products. If the enforcement of IPR laws mandates a change in how these products are designed, many such forecasting exercises will become entirely academic.
Generative AI, IPR, and Implications for India
NASSCOM’s recent report on generative AI in India states that the greatest amount of market activity related to generative AI in India is in the applications and services market. In other words, the sector reliant on using FMs created by other companies. It states that, in fact, there are no Indian FMs in existence yet. This could mean that, if copyright law is enforced in a manner that gives the authors and creators significant rights against generative AI companies and their FMs, a number of Indian companies would be significantly affected.
In addition, since most of the existing FMs have already scraped a significant portion of the information available on the internet, a lot of the value-add that FMs in India can offer is by creating more diverse and unique FMs, for example related to Indian vernacular languages, undigitized data, and so on. The process of developing these can potentially become problematic if we do not adequately consider the balance between IPR and innovation. An extreme focus on innovation would leave authors, creators, and existing knowledge repositories uncompensated and unrecognized, while an extreme focus on IPR would inhibit India’s ability to create genuine economic value and innovate.
This developing conflict between IPR and generative AI is still nascent but may have significant economic implications. In India, we must therefore initiate deliberations on this seriously to create a situation that leaves the gainers better off and no one else worse off.
—By Anirudh Burman
Review
Vito Tanzi on the Ecology of Tax Systems
Societies need governments because not all collective action problems can be solved through contractual or community-based informal arrangements. Governments need resources, and society must therefore pay for what it expects from the government. Preferably, each generation should pay for its current expenditure, while capital investments and certain kinds of current expenditures with intergenerational consequences, such as improving health, nutritional, and educational outcomes, may be incurred through borrowing, which is essentially moving resources from the future to the present. Hence, taxes and non-tax revenues, such as user charges, must be collected to pay for the current expenditure.
Arguably, the above constitute some of the first principles of fiscal policy. India’s Fiscal Responsibility and Budget Management Act, 2003, was aimed at putting some of these principles into practice, something India had never been able to do since independence. This law was made possible due to a change in context. In the 1990s, the prevailing views regarding the economic role of the government had changed. Economic reforms and favorable external conditions had helped the economy achieve moderate growth and then high growth. The government had pursued privatization and disinvestment from public sector enterprises. It was also starting to rely on the private sector’s capabilities and capital, even in sectors such as infrastructure.
However, the fiscal law has not achieved its main objectives. For instance, India has not eliminated the revenue deficit and only achieved the targeted fiscal deficit once since the law was enacted. That is also because of the context. The failure is partly on account of the economic circumstances—the economy has faced many challenges in the last two decades. Partly, it is also because a law that can be amended through money bills is not an effective restraint on a government that, by definition, enjoys a majority in the lower house of the Parliament, which is the only requisite for approving money bills.
All this suggests that the “ecology” in which a fiscal system works plays a central role in shaping its outcomes. Therefore, efforts to understand and reform fiscal policies and administrative systems must be informed by an understanding of this ecology.
In The Ecology of Tax Systems: Factors That Shape the Demand and Supply of Taxes, economist Vito Tanzi, who has worked on fiscal issues for close to six decades, offers a broad perspective on how we may think about the context for tax policy and administration. Tanzi focuses on factors such as the structure of the economy and how industrialized it is, views about the economic role of the state, the distribution of income, the openness of the economy and how globalized the world has become, the political system, the use of formal accounting, the technology of taxes (which may lead to the discovery of new taxes or to new ways of collecting taxes), the degree of decentralization and of fiscal federalism, the relative power of the national government vis-à-vis international institutions and agreements, and so on.
Tanzi shows how changes in the structure of the global economy can make it easier or more difficult to collect taxes. He shows how, on the one hand, industrialization and urbanization have progressively increased the governments’ need for public resources and also facilitated the collection of higher tax levels. For instance, they have made it easier to collect direct taxes at source from the many people who started working for larger firms. On the other hand, globalization and recent technological changes, such as the rise of global supply chains and e-commerce, have made it more difficult to collect taxes. This is especially so in the context of tax competition among countries as well as increasing tax avoidance or evasion by large enterprises and high-net-worth individuals operating globally.
Tanzi suggests that a key role in shaping tax systems is played by ideas about the role of the government, which can sometimes be dictated by necessities such as war but are often shaped by other material and ideological reasons. There is an interesting interplay that occurs between experience and ideas in a political context. Spread across several chapters of the book is a story of a rise in the role of the government, followed by a reversal that may still be ongoing. Tanzi writes about how the views on the role of the government changed, increasing the demands made from the government and legitimizing the collection of higher taxes. From the prevalent view in the pre–World War I era that the government should have a very limited role to the rise of a view that governments should have a role in redistributing incomes to the Keynesian revolution advocating for the government’s role in maintaining aggregate demand to the early development economics view that capital investment by government could boost economic growth to the human capital theory that advocated for expenditure on health, education, and so on, in just a few decades, the expectations from the government grew radically. At the same time, the economies and public administration capacities were changing in such a manner that it became easier to collect taxes to meet these demands. As a result, for several decades, the average level of taxation for many countries rose steadily.
Tanzi also shows how, beginning in the 1980s in the U.S. and UK, and then spreading to many other countries, there was a reversal of sorts. In the preceding years, the tax rates in many countries had been quite high—in some cases, the marginal tax rate was more than 90 percent. In this context, it became feasible to mobilize against the dominant views. Tanzi argues that just as the expansion of the role of government had been partly underpinned by the rising prominence of ideas that supported such a role for the government, the backlash also had its theoretical underpinnings. In chapter four, Tanzi summarizes some of them: the theory that high marginal income tax rates were implicitly subsidizing leisure, because “while the money earned from work and from additional effort was taxed, the psychic income that came from not working, that is, from remaining unemployed and idle as well as engaging in leisure activities, was not taxed”; the theory that the supply of saving could be elastic with respect to the net-of-tax rate of return that individuals get when they save, something that Keynes had dismissed; and the Laffer curve, an economic concept with questionable analytics that maintained that “tax rates above a given unspecified level might reduce, rather than increase, tax revenue.”
While these theories focused on tax rates, there were others that supported differentiating between parts of the tax base. Tanzi shows that a crucial shift that occurred was the dilution of the principle that all incomes ought to be treated alike. The view that increasingly spread was that incomes should be taxed differently depending on their source, with capital gains and interest income receiving favorable treatment. This was also enabled by globalization; the competition for increasingly mobile capital made it difficult to tax it, with some economists even advocating for zero tax on capital gains.
The combined effect of all these forces was that it became more difficult to sustain the progressivity of tax collection as marginal tax rates fell, with capital gains attracting even lower rates, and globalization made it easier to reduce the tax burden. All these developments make one wonder how much a country’s own tax policy and administration can do and how much of this challenge of tax collection is a global problem. Tanzi calls for better coordination mechanisms to “promote reform in the countries’ tax policies that have cross-country implications” as a kind of “public good at the global level.”
Another factor that Tanzi analyzes is of rising complexity in taxation. Through quantitative and qualitative evidence, he shows that tax complexity has increased in many countries. He argues that because of the incessant petitioning and lobbying in democracies, complexity tends to increase with time, even though each amendment may appear minor. Complexity makes it difficult to evaluate tax systems in terms of efficiency, equity, and certainty and increases the uncertainties in a tax system as taxpayers are unable to determine whether they are not breaking the law. Tanzi argues that this complexity also means that the theories of taxation that recommend broad schemes are not very useful. Through continuous small changes, real tax systems become far removed from their theoretical designs. So, Tanzi suggests that economists should “spend more time on the details of tax systems and perhaps less on refining theoretical and abstract designs.” He also recommends an occasional overhaul of tax laws in their entirety.
While most of the book is about the experience of developed countries, three chapters are focused on developing countries. In chapter eight, Tanzi explains why “the level of taxation in developing countries continues to be, on average, only about half of that of rich countries.” His explanation is based on certain demand-side factors (views about the economic role of government, distribution of political and financial power, citizens’ views on how well the governments use the revenues), supply-side factors (degree of urbanization, share of natural resources in exports, technology of taxes in use), and governance factors (quality of tax administration, political corruption, impact of globalization). He then analyzes the exception: some Latin American countries have been able to achieve tax-to-GDP ratios comparable with those of developed countries much earlier than expected. Tanzi shows that this has been done mainly through regressive taxes and argues that this rise in taxes has mostly not been accompanied by improvements in the quality of public services. Tanzi makes several suggestions for developing countries to overcome the obstacles to raising taxes. His main suggestion is to seek reforms in tax legislation and administration with the guiding principle of simplicity. So, he recommends avoiding high tax rates and complex deductions, minimizing the use of tax incentives, using a single, reasonable rate for the value-added tax, benchmarking the corporate tax rate to those prevailing internationally, and so on. He also emphasizes the need for international cooperation.
The book also includes a few chapters on specific aspects of the theme that read more like detours. Chapter eleven, for example, is an interesting one on how the choice of revenue-sharing arrangements in federal countries can shape the tax system; depending on the level at which the taxes are levied and collected, the outcomes may be different. There is also chapter six, which is on how the fiscal illusion created by the temporary bubble-fueled rise in tax collection in the years before the global financial crisis led to the fiscal crises subsequently. Such macroeconomic phenomena must also be considered while making fiscal policy.
Seen together, the arguments in the book present a picture of increasing challenges for tax administration, especially in democratic developing countries, where demands from the government continue to rise but meeting those demands is becoming increasingly difficult, even though a growing economy may find it easier to meet such challenges. This story is consistent with India’s experience in recent decades. India’s tax-to-GDP ratio (union and states combined) peaked at 17.9 percent in 2007–08. The government has not been able to even achieve that level since then.
While highlighting the major changes that have occurred historically, with an emphasis on the last three decades, Tanzi also cautions that the past may not be a suitable guide for the future. The ecology of tax systems is undergoing changes that may have far-reaching consequences. So, the key is to exercise careful judgment while making changes to tax policy and administration rather than falling back on standard practice or the experience of other countries. That is very useful advice for policymakers.
—By Suyash Rai