AI and digital infrastructure unite!
In the paper "Interactions Between Artificial Intelligence and Digital Public Infrastructure: Concepts, Benefits, and Challenges," Sarosh Nagar and David Eaves explore the multifaceted relationship between artificial intelligence (AI) and digital public infrastructure (DPI). They begin by defining both AI and DPI, emphasizing AI as a general-purpose technology and DPI as a specific application of digital technologies forming a foundational platform. The authors then delve into how AI can significantly enhance the public value provided by DPI through various use cases, such as language localization, fraud detection, and personalized service delivery. For example, India's Bhashini, an AI-led language translation system, aims to integrate with existing DPI to provide multilingual public services, reducing transaction costs and increasing accessibility for linguistic minorities.
Beyond AI enhancing DPI, the paper also highlights the reciprocal relationship where DPI can serve as a crucial foundation for the development of more advanced AI systems. This is primarily achieved through improvements in the quantity and quality of data available for AI training. DPI systems, by their nature, collect large amounts of consent-based data, which can help address the looming "data wall" challenge faced by large language models. Furthermore, DPI can contribute to creating standardized, high-quality datasets and reducing algorithmic bias by ensuring data from historically marginalized populations and traditional knowledge are included.
First, DPI can enhance the development of advanced Al by improving the supply of high-quality data. Currently, there are substantial fears that large language models will consume all human- generated data by 2028, with solutions like Al-generated synthetic data being imperfect substitutes due to risks that extensively training Al on synthetic data can induce the "collapse" of AI models or cause significant declines in the diversity of outputs from a model. Therefore, there is a vital need to acquire more data to enable advancements in Al systems. DPI, however, represents a values-consistent solution to this problem.
The authors also acknowledge the significant challenges in integrating AI and DPI. These include high inference costs for advanced AI models when deployed at a national scale, interoperability issues with legacy software systems used by governments, and concerns about induced bias in AI systems if DPI data is not representative. Additionally, privacy challenges related to the collection and storage of sensitive data through DPI are critical considerations. Policymakers are urged to address these technical, political, and ethical hurdles to realize the full potential of AI and DPI interactions.
The conclusions of this article are particularly relevant for the Global South and other low-resource environments. DPI, with its potential for widespread adoption and data collection, offers a unique opportunity to leapfrog traditional development hurdles. By strategically integrating AI into DPI, governments in these regions can significantly improve public service delivery, foster economic development by reducing transaction costs, and even assert digital sovereignty. The emphasis on consent-based data collection within DPI provides a valuable framework for these nations to build robust AI ecosystems that are ethically sound and representative of their diverse populations, avoiding the pitfalls of data scarcity and bias that more developed nations are now confronting.