AI as Digital Public Infrastructure
Parminder Jeet Singh
Abstract
AI is commonly regarded as a private competitive service. This essay argues that AI should instead be treated as digital public infrastructure. Constituted mainly of data and computing, AI exhibits clear public-good characteristics. Data, collectively created by people and non-copyrightable, forms a commons that requires community rights and public data infrastructures, as advanced by India’s Gopalakrishnan Committee and EU data-space initiatives. Computing is evolving into a universal utility best supplied through regulated grids. The essay contends that providing key layers of the AI stack as open public infrastructure maximises economic welfare, advances equity and social justice, and counters Big Tech monopolies. While existing global efforts demonstrate feasibility, political will for a coherent techno-legal framework is now essential.
Introduction
The phenomenon of AI is expected to herald a whole new kind of society. What parts of AI should be private, what commons, and which public, is the key political economy question. AI is basically constituted of data and computing. Data has properties as a commons. Computing in the present age should also be available as a utility like electricity. These and other factors make the case of AI as public digital infrastructure. Many initiatives exist globally where different components of AI are treated as a commons or as digital public goods. Pulling together all these into a holistic theory and practice of AI as digital public infrastructure will ensure the ‘economic welfare maximisation’ requirement, as well as the best equity and social justice in an AI society.
AI is normally taken to be a private service. Employing a new technology, a company provides answers to users’ questions or needs in exchange for payment or other commercial benefits. Multiple providers compete in this space, allowing users to switch among services. It therefore seems reasonable to view AI as a regular private good, available in a competitive market.
This paper, however, argues that AI should be treated as a public good or public infrastructure, with a basic level of access available to all as a necessity of the digital age. This is similar to how electricity and telephony have been considered in an industrial society. Such access need not be free but must be made affordable. The United Nations Development Programme (UNDP) has defined Digital Public Infrastructure (DPI) as a set of foundational digital systems that form the backbone of modern societies. AI, or at least its important components of data and computing, can and should thereby be treated as digital public infrastructure.
Whether society considers something a public good or infrastructure depends on several factors. First is the nature of the good itself: whether it is produced privately or through community means. The second is what such goods are used for, and how universally necessary they are. Everyone needs water, but not Mercedes cars. Equity considerations may therefore require a minimum level of access for all. The third factor is what has been called the “economic welfare maximisation” criterion: whether the supply of such goods as private goods maximises overall economic welfare, or whether public provision does so more effectively. While many goods can be efficiently supplied through competitive markets, others, such as electricity networks, exhibit natural-monopoly characteristics that make infrastructure-style provision more efficient. Considering them as a natural monopoly and providing them as managed utilities is generally considered the best option. And finally, practical feasibility should also be considered, even if the other necessary conditions are met.
Let us examine these conditions for public infrastructure one by one.
What Makes AI?
AI is basically constituted of data and computing. Intense, high energy-fed computing finds patterns in data to provide the AI responses and outputs. Increasingly sophisticated algorithms mediate this process with greater efficiency and accuracy. Algorithms are, of course, human creations, but once the right trajectory is established, progress can accelerate through cumulative innovation and occasional breakthroughs, such as Google’s 2017 paper on the transformer model, which was openly published and freely shared.
Data employed by AI is almost entirely created by people, whether on the Internet or other means, such as data collected from cars driven around. It is therefore difficult not to regard data as a collective good. By law, data is not copyrightable, being descriptions or statements of fact, which means data that AI companies use is a collective good. A service that finds patterns in such data, therefore, has a public-good characteristic, even if the act of finding these patterns is rewarded in some ways. The reward, however, cannot amount to exclusive control over the value derived from collectively produced knowledge, as seen among AI companies today.
Computing itself, employing immense energy, advanced chips, and algorithms, can arguably be considered a private activity. However, it continues to draw energy from public sources. Also, as computing becomes increasingly commoditised and essential to social and economic life, it begins to resemble a crucial utility, best provided as a universal infrastructure for all with equity-based, regulated charges.
Much as electricity is made available as a managed utility because of its universal importance, computing power is increasingly required as a cognitive adjunct to human activities at work and at home. As computing needs grow rapidly, the case strengthens for treating basic computing infrastructure as a shared utility rather than a fragmented collection of captive systems. Such an approach would meet the economic welfare maximisation criterion and reduce computing equipment and power wastage.
A few words might already be said here about AI itself as public infrastructure, before turning to this issue in greater detail. Unlike electricity, where only the output matters, both input and output matter in the case of AI. AI systems should reflect the populations and conditions they serve. For this, any kind of universal or specific AI must be fully representative of the groups it serves. Only a public interest-oriented system can appropriately ensure this.
Second, AI is increasingly becoming a near-universal cognitive tool integrated into everyday life, both at work and at home. AI itself, and not just its main components of data and computing, thus meets the key criteria to be classified as a digital public infrastructure.
Data as Infrastructure
If data comes from everyone, it follows that everyone should have access to it as infrastructure. Data rights have almost exclusively been discussed and legislated at the individual level. While this provides privacy safeguards for individuals, AI systems rely primarily on collective data. There currently exist no legal provisions governing rights over such collective data. As a result, those who collect data become the de facto owners, with no obligations to the communities that contributed the data.
A Government of India Committee of Data Experts under Kris Gopalakrishnan addressed this issue head-on. Its June 2021 Report introduces the concept of “community data” to describe collective data collected from various communities, and gives the community concerned legal ownership and rights over the corresponding collective data. It also developed a draft law for such community ownership of data and its effective use, and the entities that the respective communities would like to be able to employ such data for their benefit. (It could be a community’s commuting data collected by Uber to be used by various community-level applications through local start-ups.) The report also outlined an elaborate architecture for the implementation of its recommendations. Under this model, the relevant community would be represented by community data trusts. All collectors of data considered to be of high public value would have to contribute the data to these trusts, which would then run data infrastructures to provide the data to eligible users – including start-ups, research organisations, public purpose entities, and others.
Similar data infrastructures have also been mooted in French government documents. The cities of Hamburg and Mainz in Germany are developing municipal data utilities on the premise that the collective data produced by city dwellers and otherwise belongs to the city. However, these propositions have not yet been taken forward to develop full-fledged, across-the-sectors, actual data infrastructures to supply data for community use, through mediating local businesses, non-governmental organisations, and research entities.
Similarly, the European Health Data Space Regulation mandates a data infrastructure for personal health records, as well as for collective health data for policy, innovation and research purposes. The deadline for its implementation, however, is still a few years away.
The EU also has public data infrastructures in the area of transport and tourism, although these are not accompanied by legal mandates requiring data sharing by all relevant collectors of public data. These initiatives nevertheless help establish the practical and technical foundations upon which broader public data infrastructures can be built. Once legal mandates and Europe-wide public data systems exist, these may be used by local AI companies, especially smaller ones without access to sectoral or general data. The EU Data Strategy lays out plans for Data Spaces across all important sectors. The Health Data Space Regulation may represent the first step towards legal mandates for sector-wise data infrastructures.
The EU’s GAIA project, sponsored by the German and French governments, provides another interesting example. It takes a bottom-up approach to bring together willing actors in different sectors to develop common data spaces and data infrastructures. The term GAIA-X is often used, with X standing for different sectors like agriculture, tourism, education, and the auto industry. Some communities in the EU, even outside the GAIA project, and in other places worldwide, have been experimenting with cooperative agriculture data trusts. While GAIA has seen some success, such bottom-up efforts need to be complemented by government-mandated data infrastructures, including legal mechanisms requiring the contribution of data to shared public systems.
Bringing these promising initiatives together requires political will of a very high order, which even the EU has not been able to muster. Its governments and local businesses remain divided on whether to bet on such an alternative public digital infrastructure approach to digital development and digital sovereignty, or to try to replicate the U.S. private hegemonic models. This hesitation risks leaving them caught between competing approaches. The winner-takes-all U.S. model already has entrenched winners, while the EU risks losing valuable time in developing an alternative model of private digital enterprise built on public digital infrastructure.
Meanwhile, the rest of the world waits for the EU’s policy direction and the outcomes of its still-piecemeal efforts before taking any decisive steps of their own. The implicit mindset is: if the EU itself cannot do it, others are unlikely to succeed. This has resulted in a stalemate that hinders progress towards a public digital infrastructure-centred digital model at scale. Yet such a model may offer one of the few viable alternatives to the current, hegemonic Big Tech-centred digital economy, universally considered both unsustainable and dangerous.
The necessary elements to develop country- or region-wide data infrastructures are now visible, and practically at hand. The EU’s Health Data Space Regulation has strong provisions for mandated data contribution and practical mechanisms for sharing data with eligible users. India’s Gopalakrishanan Committee has a draft law in place, as well as the blueprint for legal/organisational (community data trusts) and technical structures of actual pipes and control centres for data distribution. The EU also has elaborate data infrastructures in areas like transport. GAIA and other such projects worldwide have made significant bottom-up progress in these directions.
An elaborate theory, law, and practice of country- or region-wide sectoral data spaces must now be developed and enforced as the only means of challenging Big Tech’s data and AI power. The most valuable data is unlikely to be shared voluntarily, data being the new oil or gold. It is necessary to legally recognise collective data as a collective resource over which communities have rights of access, use, and governance. The fruitless debate between voluntary contribution and mandatory data sharing may take one nowhere. No more than a pharmaceutical company would voluntarily share its key IP (intellectual property), will Big Tech voluntarily share its most valuable datasets. It is wise to remember that Big Pharma, and the countries supporting it, fought huge battles (including at the WTO) not to share IP for COVID-critical medicines during a global health emergency. On what basis, then, does one expect Big Tech to really ever share their most valuable data?
A data sharing law is therefore the starting point for meaningful and effective public data infrastructures, or data as a digital public infrastructure. Public and community bodies, under a dedicated regulator, can then undertake the implementation work of building the infrastructure required for data distribution.
Computing as Infrastructure
As computing needs become more demanding, and our devices smaller, most computing is already done remotely. Cloud computing now underpins not only our work and home lives, but also industrial units and service companies. Last year, a major oil company in India found itself suddenly locked out of all its digital systems, including its own data, because the U.S. company providing these services responded to EU sanctions. This is an untenable situation, especially as more and more activities depend on remote computing, much of it currently foreign-based.
Cloud computing has two elements: raw computing power and remotely hosted applications and data. It is increasingly considered best practice to host the actual data as close to the user as possible, preferably in the same country. These data centres must be strictly regulated so that companies cannot be denied access to their own data. Applications, even foreign ones, can similarly be hosted locally, with appropriate regulatory controls to ensure continuity of services, much like for utility companies. The same principle should apply to computing hardware, with alternative computing power available over a utility-like grid.
Computing power grids have been employed by the EU and some countries, even if with limited scope or purposes, such as for educational and research facilities. High Performance Computing (HPC) projects have existed in the EU for a long time now, supporting education, research, and innovation, including for small enterprises. The previously mentioned GAIA project has been developing a “European cloud,” networking different cloud providers with common standards and strong service conditions for reliability, interoperability, and easy switching.
Dependence on captive computing power makes it difficult for digital markets, especially AI, to become genuinely competitive, thus resulting in an oligarchy, with two digital superpowers ruling the world. Reliance on Big Tech’s computing power creates not only reliability concerns but also other innovation-suppressing lock-ins. While computing power is not in itself a source or site of innovation, deprivation or highly conditioned access to it may impede innovation. A more competitive digital ecosystem requires computing resources to be broadly available through well-regulated infrastructure, enabling a wider range of firms to develop digital and AI products.
India recently developed a GPU (high-end chips needed for AI) grid and is making it available to Indian AI companies. This model has to be expanded into a nationwide High-Performance Computing digital public infrastructure of the kind that the EU has been developing.
All this strengthens the case for treating computing as digital public infrastructure, alongside data, the other key component of AI.
Infrastructural Elements of the Computing/AI Stack
When IBM made its first office computers, it was a fully integrated system. Hardware, software, computing, and data were bundled together. In a hurry to enter the personal computing market, IBM developed a personal computer and asked Microsoft to provide its software. Microsoft, instead of selling it to IBM, licensed it, and thus retained the right to license it to other personal computer makers. With this separation of hardware from the software layer, a software revolution began. Later, when the Internet began to gain widespread adoption, public interest engineers behind its development created the idea of layered digital architecture, whereby each layer could operate independently of other layers, and would remain open to competing digital possibilities. From here on, the world began to see computing as a stack of separate layers of digital artefacts. This layered approach is similar to how infrastructures have traditionally been understood: as separate, self-sufficient layers that are equally open and available to different activities and possibilities. This goes against the vertically integrated physical products and systems typical of the industrial age, such as manufactured goods and service enterprises.
AI companies nowadays, however, favour full vertical integration, from chips to AI models, to stave off competition, especially from smaller players. If this process continues unchecked, there may be little prospect of a competitive AI field or of a country other than the two current AI superpowers developing significant capabilities. During the Gilded Age of corporate concentration and corruption in the U.S. of the late nineteenth century, regulators stepped in to break the vertical integration of mining and manufacturing with railroads and highways. Similar interventions later targeted oil companies and, subsequently, the vertical integration of telcos. Such action may now be needed for Big Tech digital corporations to break the vertical integration across digital layers, from hardware and cloud computing to data, software/ algorithm and AI models. This author argued for such horizontal separations in a 2020 paper, Breaking Up Big Tech: Separation of its Data, Cloud and Intelligence Layers.
The challenge is that these regulatory breakups took place in the U.S. at a time when national political institutions intervened in national and public interest. Today, however, vertical AI integration is concentrated in U.S. and China, while many of its costs are borne elsewhere, even as both countries derive significant strategic power from it. This makes an effective political response to the issue considerably more difficult. The EU has taken measures like its Digital Markets Act that seek to enforce some such structural separations but has not implemented it with the strength and clarity that is required to see substantial results. Meanwhile, the U.S., in its discussions with the EU, keeps threatening the EU against implementing this Act against the U.S.’s Big Tech, as it does to any other country attempting any digital competition measures.
What is needed, therefore, is a global movement to structurally separate the key digital layers. In the language of digital activists, this would make these layers more “open.” The more important of these digital layers need to be provided as digital public infrastructure. One of the tragedies of the digital rights movements has been that the “openness movement” led by digital activists was unable to build stronger connections with traditional public infrastructure thinking. This has to do with the individualist, state-suspicious mindset of digital activists, and therefore of the movement itself. Yet country/region-wide digital public infrastructures require an active role for governments, albeit within frameworks governed by appropriate rights-based laws and independent regulators. Traditional infrastructure actors must also recognise the new and significant governance challenges posed by developing digital public infrastructures.
There is an urgent need for “openness” activists to recognise the role the state can play in advancing many of the goals they have long championed. Equally, advocates of public digital infrastructures must recognise activists’ concerns that such systems can deepen state intervention in people’s rights and freedoms. Both groups should work towards building common ground on the rights-based laws and safeguards needed to protect privacy and freedoms.
Once these chasms are somewhat bridged, digital activists and those involved in economic and social rights, as well as national development, need to come together to develop a powerful movement for digital public infrastructure in all digital/AI layers, including the ones discussed in this paper. Debates over possible state excesses, on the one hand, and private greed, on the other, will persist in democratic spaces. Yet such a coalition may be the only way to challenge the unsustainable AI power of Big Tech, which threatens to destroy not just equity and justice, but our very civilisation.
Chips, as the basic hardware for AI, need to be structurally separated in production and distribution from higher layers, much as computer hardware and its components were separated from higher software layers in an earlier era. Strict regulation is needed to ensure the openness of these chips and hardware layers. Open source software already powers many of the digital processes today. New regulations may be needed surrounding its sharing, including modifications made to it. The same applies, to the extent possible, to AI research and models. With these layers open, innovation can focus on competing to provide the best services on top of them. This would likely be most economic welfare maximising and more equitable.
Today, even public-interest-minded researchers often have to join Big Tech to conduct their research, because sectoral data (like with Uber for transport) and all the huge computing power remains concentrated within a few firms. This is unlikely to be sustainable in the long run. Much of the most important basic research has always been conducted in universities and independent research centres and then made widely available in the public interest. This further underlines the need to open up data and computing layers and provide them as digital public infrastructure.
AI as Digital Public Infrastructure
All the AI around us is built on collectively produced human knowledge. GenAI, for instance, learns by identifying patterns in data and associating them with desirable outcomes, including recognising linguistic structures through relationships among words and phrases. At its core, it involves algorithms operating on collective human knowledge. While a company making good algorithms for this purpose would need to be rewarded appropriately and proportionately, such rewards should not amount to full ownership over the value derived from human knowledge.
The above is the digital public infrastructure rationale for AI from the perspective of what goes into making AI. Second, AI is increasingly becoming an elementary need as cognitive tools are embedded into all or most human activity. Its increasing necessity and universality strengthen the case for treating AI as a digital public infrastructure. The logic of economic welfare maximisation, aiming to avoid wasteful duplications, as well as that of equity, becomes self-evident here.
When OpenAI was founded, Sam Altman described AI as a universal necessity which could be provided as a utility, like water and electricity. Elon Musk expressed similar views during the early years. Subsequent developments have arguably strengthened this logic, regardless of how the positions of these early proponents may have shifted towards personal greed and the need for extraordinary monopoly powers for themselves and their country. The times since have only strengthened this logic.
Yoshua Bengio, considered a godfather of AI, meanwhile, has retained his focus on AI as a shared civilisational heritage, and its availability as a common resource. He is now working on separating the knowledge-generating component of AI, which he calls “Scientist AI,” from the agentic part that applies this knowledge to particular tasks as desired by AI’s specific human handlers. The aim is for the core, non-agentic part of AI to remain available as a common resource, or a digital public infrastructure. This would both check the excesses of agentic systems and develop safer agentic systems on top of the core “Scientist AI.”
Some Practical Considerations
What is needed is a full-scale techno-legal approach, as iSPIRT, the Indian volunteers’ group behind India’s digital public infrastructures movement, calls it. Too often, like in the EU, half-hearted purely legal approaches are employed, while elsewhere, like in India, the emphasis is placed on technical solutions. Neither approach is sufficient on its own. A holistic approach, spanning theory and technical and regulatory practice, needs to be developed and implemented.
AI appears to be evolving with such force that it has led to complete policy paralysis. First, we need to recognise that while AI may be a transformative technology, societies have successfully developed policy responses to major technological shifts in the past. The fact that it is a global phenomenon, and the home countries of Big Tech prioritise promoting them as a means of national power over regulation, undoubtedly complicates matters. The proper response in this regard is for other governments, research organisations, and non-profits to collaborate in developing a full-fledged theory and practice of AI as a digital public infrastructure, with key layers of the AI stack remaining fully open.
As this paper has suggested, all the basic concepts and pieces of practice are already in place. With sufficient political will, these have to be pulled together into a coherent legal and workable structure. With practical considerations in mind, especially the rapid uptake of existing AI, this may be a gradual process. At the same time, it has to be a completely clear and purposeful direction. Policymakers should clearly articulate whether a public digital infrastructure approach, with open digital layers, is to be the future of AI. When the political will is unmistakably evident, all players will begin adapting to that politically determined future. Battles over standards, international law, and trade policies will still have to be waged, but they are easier to address with a coherent, holistic approach that is already laid out.
We have discussed many examples of regulated computing infrastructures that can be made available across society and layered for different kinds and levels of users. This, combined with a local system of well-regulated data centres and infrastructures, and utilities like strict service conditions, can allow for the creation of a “computing as digital public infrastructure” system.
We have seen how the Gopalakrishnan Committee laid out a draft law for community data and its use, along with the technical and regulatory architecture needed to support such public data infrastructure. Similar practices have been undertaken in the EU and in some Western cities.
iSPIRT has developed an interesting concept of public digital infrastructure for interaction between data and AI models, known as “confidential clean rooms.” These are technical systems operated by independent third parties (like Account Aggregators in India’s banking system) that enable independently assured interactions between AI models and data. AI models can abstract learnings from such data while adhering to parameters and conditions specified by data providers, which are ensured by a third party. The EU’s Data Governance Act similarly provides for third-party data repositories and providers certified by regulators and operating on public interest principles.
Similar third-party digital public infrastructures can be developed and operated for various layers of the AI stack, under the supervision of dedicated regulators, to form a viable “AI as digital public infrastructure” model. To overcome the current policy and political-will paralysis, policymakers and other stakeholders must first be able to clearly and unmistakably envision a viable, realistic, fully workable, and scaled-up “AI as digital public infrastructure” model. The absence of such a clear vision contributes to hesitant policy responses. Scholars, public-interest advocates, research centres, and civil society groups should therefore focus on developing a coherent vision for policymakers that brings together the many elements of this complex puzzle in a realistic manner. This paper is a brief, humble attempt in that direction.