The role of artificial intelligence in achieving the Sustainable Development Goals

The emergence of artificial intelligence (AI) and its progressively wider impact on many sectors requires an assessment of its effect on the achievement of the Sustainable Development Goals. Using a consensus-based expert elicitation process, we find that AI can enable the accomplishment of 134 targets across all the goals, but it may also inhibit 59 targets. However, current research foci overlook important aspects. The fast development of AI needs to be supported by the necessary regulatory insight and oversight for AI-based technologies to enable sustainable development. Failure to do so could result in gaps in transparency, safety, and ethical standards.

The emergence of artificial intelligence (AI) is shaping an increasing range of sectors. For instance AI is expected to affect global productivity 3 , equality and inclusion 4 , environmental outcomes 5 and several other areas, both in the short and long term 6 . Reported potential impacts of AI indicate both positive 7 and negative 8 impacts on sustainable development. However, to date there is no published study systematically assessing the extent to which AI might impact all aspects of sustainable developmentdefined in this study as the 17 interconnected Sustainable Development Goals (SDGs) and 169 targets internationally agreed in the 2030 Agenda for Sustainable Development 1 . This is a critical research gap, since we find that AI may influence the ability to meet all Sustainable Development Goals (see a summary of the results in Fig. 1, and full results in the Supplementary Table 1).
Here we present and discuss implications of how AI can either enable or inhibit the delivery of all 17 Goals and 169 targets recognized in the 2030 Agenda for Sustainable Development. Relationships were characterized by the methods reported at the end of this article, which can be summarized as a consensusbased expert elicitation process, informed by previous studies aimed at mapping SDGs interlinkages 9 .For this study, we adopt Russell and Norvig's definition of AI as a field that "attempts not just to understand but to build intelligent entities" 2 (see full definition in the Methods section). This view encompasses a large variety of subfields, including machine learning.

Documented connections between AI and the SDGs
Our review of relevant evidence shows that AI may act as an enabler on 128 targets (76%) across all SDGs, generally through a technological improvement which may allow to overcome certain present limitations. However, 58 targets (34%, also across all SDGs) may experience a negative impact from the development of AI. For the purpose of this study, we divide the SDGs into three categories, according to the three pillars of sustainable development, namely Society, Economy and Environment 10,11 (see the Methods section). This classification allows us to provide an overview of the general areas of influence of AI.  Substantive research and application of AI technologies to SDGs is concerned with measuring or predicting certain events using, for example, data mining and machine-learning techniques. This is the case of applications such as forecasting extreme weather events or predicting recidivist offender behavior. The expectation with this research is to allow the preparation and response for a wide range of events. However, there is a research gap in real-world applications of such systems, e.g. by governments (as discussed above). Barriers for institutions for adopting AI systems and data as part of their decisionmaking process include the possibility to adopt such technology while protecting the privacy of citizens and data, high cyber-security needs, and the technical capabilities needed to have AI-systems functioning properly. Targeting these gaps would be essential to ensure the usability and practicality of AI technologies for governments. This would also be a prerequisite for understanding long-term impacts of AI regarding its potential, while regulating its use to reduce the possible bias that can be inherent to AI 8 .
Furthermore, our research suggests that AI applications are currently biased towards SDG issues that are mainly relevant to those nations where most AI researchers live and work. For instance, many systems applying AI technologies to agriculture, e.g. to automate harvesting or optimize its timing, are located within wealthy Western nations. Our literature search resulted in only a handful of examples where AI technologies are applied to SDGs-related issues in nations without strong AI research. Moreover, if AI technologies are designed and developed for technologically advanced settings, they have the potential to exacerbate problems in less wealthy nations (e.g. when it comes to food production). This finding leads to a substantial concern that developments in AI technologies could increase inequalities between wealthy and less wealthy nations, in ways which counteract the overall purpose of the SDGs. We encourage researchers and funders to focus more on designing and developing AI solutions which respond to localized problems in less wealthy nations. Projects undertaking such work should ensure that solutions are not simply transferred from a wealthy nation. Instead, they should be developed based on a deep understanding of the respective region or culture to increase the likelihood of adoption and success.

Towards Sustainable AI
Our assessment of published evidence shows that AI can have a positive impact on all the SDGs. This is essentially through technological breakthroughs that will lead to better outcomes in several sectors. However, there are a number of problems associated with AI that if not addressed may inhibit the achievement of several SDGs.
First, the great wealth that AI-powered technology has the potential to create may go mainly to those already well-off and educated, while job displacement leaves others worse off. Globally, the growing economic importance of AI may result in increased inequalities due to the unevenly distributed educational and computing resources throughout the world. Furthermore, the existing biases in the data used to train AI algorithms may result in the exacerbation of those biases, eventually leading to increased discrimination. Other related problems are the political polarization due to the massive use of social media, the lack of robust research methods to assess the long-term impact of AI, and privacy issues related to the data-intensiveness of AI applications. Many of these aspects result from the interplay between technology development, requests from individuals and response from governments. Figure 5 shows a schematic representation of these dynamics.  It is also imperative to develop regulations regarding transparency and accountability of AI, as well as to decide the ethical standards to which AI-based technology should be subjected to. This debate is being pushed forward by initiatives such as the IEEE ethical aligned design 45 , and the new EU ethical guidelines for trustworthy AI 46 .In this sense, the lack of interpretability of AI, which is currently one of the challenges of AI research, adds an additional complication to the enforcement of such regulatory actions 47 . This, however, implies that AI algorithms, which are trained with data consisting of previous regulations and decisions, may act as a "mirror" reflecting biases and unfair policy. This presents an opportunity to possibly identify and correct certain errors in the existing procedures. Again, the friction between the uptake of data-driven AI applications and the need of protecting individuals´ privacy and security is stark. When not properly regulated, the vast amount of data produced by the citizens might potentially be used to influence consumer opinion towards a certain product or political cause 48 .
We are at a critical turning point for the future of AI. A global and science-driven debate to develop shared principles and regulations among nations and cultures is necessary to shape a future in which AI positively contributes to the achievement of all the Sustainable Development Goals. All actors in all nations should be represented in this dialogue, to ensure that no one is left behind. On the other hand, postponing or not having such conversation could result in an unequal and unsustainable AI-fueled development.

Methods
In this section we describe the process employed to obtain the results described in the present study and shown in the Supplementary Table 1. The goal was to answer the question "Is there published evidence of AI acting as an enabler or an inhibitor for this particular target?", for each of the 169 targets within the 17 SDGs. To this end, we conducted a consensus-based expert elicitation process, as discussed by Butler et al. (2015) 49 and Morgan (2014) 50 . The authors of this paper are academics spanning a wide range of disciplines, including engineering, natural and social sciences, and acted as experts for the elicitation process. The authors performed an expert-driven literature search to support the identified connections between AI and the various targets, where the following sources of information were considered as acceptable evidence: • Published evidence on real-world applications (given the quality variation depending on the venue, we ensured that the publications considered in the analysis were of sufficient quality). • Published evidence on controlled/laboratory scenarios (given the quality variation depending on the venue, we ensured that the publications considered in the analysis were of sufficient quality). • Reports from accredited organizations (for instance: UN or government bodies). • Documented commercial-stage applications. On the other hand, the following sources of information were not considered as acceptable evidence: • Educated conjectures, real-world applications without peer-reviewed research.
• Other sources of information.
We considered any software technology with at least one of the following capabilities as relevant: perception -including audio, visual, textual, and tactile (e.g. face recognition), decision-making (e.g. medical diagnosis systems), prediction (e.g. weather forecast), automatic knowledge extraction and pattern recognition from data (e.g. discovery of fake news circles in social media), interactive communication (e.g. social robots or chat bots) and logical reasoning (e.g. theory development from premises). The list of connections between AI and the 169 targets, together with a paragraph summarizing the reasoning behind the assessment of AI acting as a potential enabler or inhibitor of that particular target, is available in the Supplementary Table 1. A list of references supporting the reasoning is also provided for each of the targets.
The expert elicitation process was conducted as follows: each of the SDGs was assigned to one or more main contributors, and in some cases to several additional contributors as summarized in Table 1. The main contributors carried out a first literature search for that SDG, and assessed whether the published evidence reflected positive or negative impacts of AI on the various targets of that particular SDG. Then the additional contributors (if assigned to that SDG) completed the main analysis with additional references and discussions with the main contributors. One published study on a synergy or a trade-off between a target and AI was considered enough for mapping the interlinkage. However, for nearly all targets several references are provided. After the analysis of a certain SDG was concluded by the contributors, a reviewer was assigned to evaluate the connections and reasoning presented by the contributors. The reviewer was not part of the first analysis, and we tried to assign the roles of main contributor and reviewer to experts with complementary competences for each of the SDGs. The role of the reviewer was to bring up additional points of view and considerations, while critically assessing the analysis. Then main contributors and reviewers iteratively discussed to improve the results presented for each of the SDGs. This process was conducted through regular meetings over approximately six months, until the analysis for all the SDGs was sufficiently refined.
After reaching consensus regarding the assessment shown in Supplementary Table 1, we analyzed the results quantitatively by evaluating the number of targets for which AI may act as an enabler or an inhibitor. A total of 128 targets reflected positive impact of AI, whereas for 58 the literature indicated negative impact. This corresponds to 75.7% and 34.3% of targets with positive and negative impact, respectively. Furthermore, we carried out the same analysis for each of the SDGs, and calculated the percentage of targets with positive and negative impact of AI for each of the 17 Goals, as shown in Fig. 1. Additionally, we divided the SDGs into the three following categories: Society, Economy and Environment, consistent with the classification discussed by Refs. 10,11 . The SDGs assigned to each of the categories are shown in Fig. 6 and the individual results from each of these groups can be observed in Figs. 2-4. These figures indicate, for each target within each SDG, whether any published evidence of positive or negative impact was found.