Demystifying the AI Monolith for Sustainable Development Part 1: Observations on Our Use of Data and Computing

By Jeremy Boy

May 29, 2024

Farmers in a field with computer image spreading throughout — Farmers of the future. Created using Midjourney.

Artificial Intelligence (AI) is trendy, but hard to grasp. It is often discussed as a monolithic technology with the promise of global societal impact. Other times it is discussed (and debated) as a more conceptual undertaking of trying to reproduce the human intellect. However, in many applications it is still a simple—albeit advanced—set of computational techniques that are useful for clustering and classifying data. Whatever the lens, the conversation on AI in the context of sustainable development tends to portray it as a double-edged sword: on the one hand, AI is believed to hold the potential to accelerate progress towards the Sustainable Development Goals (SDGs); on the other, it is feared to cause harm to underrepresented or disenfranchised communities.

This blog post is intended to be the first in a series in which we will try to unpack these conceptions of AI, by looking into the work of the UNDP Accelerator Labs Network. Here, we begin by investigating how 19 Labs are experimenting with a combination of unusual data sources and advanced computational techniques to help address different development priorities around the world. We focus on AI from the perspective of data innovation and look at concrete Labs’ experiments by referring to the following triptych:

Data sources
Processed via computational techniques
Used to better act on development priorities (aka application areas).

Data innovation in the Accelerator Labs Network

Over the past four and half years, the UNDP Accelerator Labs Network has been actively mainstreaming data innovation inside and outside UNDP, with the explicit goal of diversifying the ways in which we understand sustainable development problems and identify solutions to those problems. Take for example, our work on collective intelligence, which focuses on connecting people’s knowledge with data and technology to generate new insights and action (check out our recent publication with Nesta’s Centre for Collective Intelligence Design on the topic of climate action); or our strong presence in UNDP’s Make AI Work for Us feature, and in the Data to Policy Navigator.

Here, we take stock of these efforts in a meta-analysis of 27 experiments conducted by 19 Labs from all regions of the world. In most cases, these experiments showcase a relatively advanced level of technical maturity and have been tested in real world situations. Our aim is to start understanding what data with what technique are best suited for what sustainable development priority, hence the triptych.

The data innovation triptych

Fig. 1 shows how each experiment flows through the data source—computational technique—application area triptych.

A graph showing the triptych of data innovation work — Fig. 1: Unpacking the data innovation work of the Accelerator Labs Network. Browse the interactive version of the diagram here: https://undp-accelerator-labs.github.io/AI-map/

Let’s start with the data source. At a high level, we see that Earth observation data, whether low(er) quality satellite imagery or high(er)-resolution drone imagery, is the most used in our sample of experiments. This is likely due in part to the relative simplicity of accessing imagery, and the maturity of associated computational techniques.

We then see a trend in how these data are used, and what they are used for, in other terms the “application areas.” Precision agriculture and pastoralism, as well as land, water and air pollution and waste management are the most common application areas across Labs, often aiming to detect and classify features in the imagery, like the aggregation of solid waste or crop diseases, or for deriving land use/land cover maps and vegetation indices.

For example, the UNDP Accelerator Lab in Cabo Verde, in partnership with the University of Cabo Verde, Prime Botics, and the Associação de Produtores da Ribeira de São Filipe, has developed an integrated system that uses drones equipped with cameras and specifically trained computer vision algorithms to identify pest infestations in bean crops. Similarly, the UNDP Accelerator Lab in Cameroon has worked on a drone imagery-based computer vision technique to facilitate the detection of crop diseases in cocoa trees.

Men watch as a drone takes off — The UNDP technical team in Cameroon prepares for the launch of an experimental drone for a local cocoa planter to help monitor the growth of his cocoa trees.

Photo: UNDP/Rodrigue Martial Mbarge et Severin Ndemanou

Meanwhile, the UNDP Accelerator Labs in Bolivia, Guatemala, India, the Philippines, Serbia, and Viet Nam have explored the use of satellite imagery for detecting features related to pollution. The Lab in India has focused on the construction industry that generates air pollution; the Labs in Guatemala, the Philippines, Serbia, and Viet Nam have focused on the accumulation of solid waste; and in Bolivia, the Accelerator Lab has focused on environmental liabilities related to mining.

We also see a small trend across Labs that use online posts, typically from travel sites, to understand sub-national tourism dynamics. For example, the UNDP Accelerator Lab in Tanzania, in partnership with the SDG AI Lab and the Zanzibar Commission for Tourism, is building an analytics tool that pulls data from popular tourism websites to generate new insights on visitor dynamics on the island of Zanzibar, and to help create new travel experiences for visitors. It consolidates lists of popular attractions in Zanzibar City, such as museums and historical sites; visitor feedback on the different attractions; aggregated and anonymized demographic information on visitors of the different attractions; temporal dynamics of the visits; and the overall sentiment of visitors. The tool builds on the success of a previous experiment the Lab conducted with Xsense AI. Similarly, the UNDP Accelerator Labs in Malawi and Jordan have explored the use of social media monitoring to better understand sub-national tourism dynamics.

Another notable pattern is the fact that the Network is experimenting with generative AI, testing both image and text-based applications for very diverse sustainable development priorities, ranging from urban planning to political discourse analysis. For example, the UNDP Accelerator Lab in North Macedonia, in partnership with Urbanist AI, has designed a novel methodology for conducting public consultations on urban planning. Participants are offered a hands-on experience with an AI-powered urban image generator tool, with which they are invited to create visual proposals for redesigning public spaces. The Lab tested this methodology during a one-day workshop in Skopje that focused on revitalizing public spaces that particularly suffer from micro-heat island effects. People from diverse backgrounds, including local government, civil society, architecture students and urban planning enthusiasts, used the tool to collectively imagine a greener, cooler landscape for their city. The workshop resulted in a wealth of creative options for three distinct public spaces in Skopje.

In addition to North Macedonia, the UNDP Accelerator Labs in Argentina, Guatemala and Cameroon have also begun exploring the potential of generative AI for sustainable development. We see this as a sign of the Accelerator Labs Network’s research and development (R&D) capacity, as Labs are able to quickly react to the development of new technologies and test their potential for contributing to sustainable development priorities. (For more on the UNDP Accelerator Labs’ take on R&D for sustainable development, you can read our secret plan and manifesto.)

What’s next in this series?

It is important to note that all these experiments can fit under the banner of AI for sustainable development, despite the diversity of ways in which they flow through the data innovation triptych. Bringing them to scale further involves and impacts very diverse stakeholders, from farmers to waste pickers to government agents. Our sample also likely does not paint the whole picture, even within the UNDP Accelerator Labs Network. This should force consideration of how abstracting AI to a monolithic technology can be problematic when discussing the ethics of its applications or formulating best-practice or regulatory guidelines and policies.

Moreover, these experiments apply AI in complex socio-technical systems as a means to drive action towards the SDGs. This contrasts with more administrative or evaluative applications, like classifying reports and developing new ways of measuring progress on given development indicators. As such, while the technology is generally mature (the tech works!), it is important to stress that not all implementations in real-world contexts will necessarily be successful. Indeed, it is not because the technology is ready that the social system in which it is injected is, or that the effects of the technology that initially seemed desirable end up truly being so. For example, introducing precision agriculture techniques to an environment primarily composed of smallholder farmers not only begs the question of how farmers might afford the necessary technology, but also forces a wider re-consideration of how they interact with and work their land. There are undoubtably strong anthropologic dimensions to these emerging factors (see points 12–14 in How Complex Systems Fail).

In the next blog of this series, we will focus on unpacking these more systemic implications to better understand what makes for success or failure in the deployment of AI for sustainable development, beyond technical feasibility. We will explore the human and social factors that impact the adoption of the proofs of concept that have been developed for these experiments. We will also keep updating our interactive version of the diagram shown in Fig. 1 on GitHub, with more UNDP Accelerator Labs' work, so keep an eye on it!

Methodological note: The Accelerator Labs “work out loud,” meaning they continuously publish updates on their work, whether through blogs or action learning plans and written reflections. We sampled the 27 experiments from 19 different Labs using these sources of information, looking for terms like “data innovation,” “unusual data,” “big data,” “data science,” “machine learning,” “artificial intelligence” and “AI.” Get in touch to learn more about our methodologies at accelerator.labs@undp.org.