Michaela Wenner

Artificial intelligence (AI) and big data create possibilities in sustainable development research and real-life applications. Inherited biases and unequal resource and education capacities, however, threaten to increase inequality in many areas and negatively influence parts of the sustainable development goals.

Recently, Vinuesa et al. (2020), published a systematic evaluation of AI on sustainable development goals (SDGs). For a majority of targets, AI promises to have a positive effect. In contrast, some target accomplishments are threatened by this emergent technology. Figure 1 (Vinuesa et al., 2020) shows a summary of the effect of AI on different sectors of the SDGs. As the primary trend, one can identify that positive impacts dominate in all areas. For some goals, however, adverse effects are substantial. Interestingly, AI seems to have a significant positive impact on the SDG of no poverty (SDG 1) and quality of education (SDG 4), but it has the most substantial negative consequences.

Figure 98.1 – Positive (a) and negative (b) impacts of AI on SDGs (Vinuesa et al., 2020). The outer circle shows three different categories of targets with the percentages showing the proportion of potentially affected targets by AI. The inner circle shows the effect on the individual goals. The number in brackets and the inner shaded areas show the percentages when the type of evidence is taken into account.

The positives. Technological improvements and increasing computational power enable a broader usage of AI in many fields. For SDGs, this might be connected to the detection of poverty areas, the understanding of climate change, creating a smart city, or efficient electricity usage. Additionally, it might help to provide basic human needs, such as food and clean water. Especially in the health sector, AI has proven as a reliable tool to identify an illness. From an economic point of view, increased productivity can be attributed to the emergence of AI. Natural sciences can also make use of AI, e.g., biodiversity and natural hazard monitoring or the detection of oil spills. However, many of these positives carry a large package of negative aspects, ranging from carbon emissions due to computational power, to across border inequality, as well as bias on a national basis. In the following, the central (potential) negative aspects identified by Vinuesa et al. 2020 will be discussed with short examples.

Bias. Many AI algorithms require a training data set, which often has been labeled by humans or drawn from historical data. This introduces an inherent bias in the training data set toward minorities and gender. Risk-assessment algorithms in the US, for example, have been shown to be biased towards minority and low-income communities (2). An increased target of such communities by law enforcement will increase the number of detected crimes and therefore lead to a high recidivism rate. A high recidivism rate, in turn, leads to more law enforcement being sent to such communities. This self-enforcing loop, starting at a biased training data set, results in a much larger bias towards minorities. Similar patterns can be observed in gender bias, such as in natural language processing (3). So-called word embeddings, a popular language processing technique, tend to link, for example, technical jobs more with men than with women. On the other hand, domestic activities are more linked to women than men.

A different kind of bias results from the self-interest of publishers to only publish positive results. The community is more likely to investigate the enabler of SDGs than inhibitors. Therefore, a systematic assessment can be biased towards more positive impacts, only due to the addressed research topics. This makes it hard to get a broader picture of the impact of AI on SGDs, as scientific gaps exist, and long-term studies are rare in the young field of AI research. Additionally, AI research is often funded by industry, which leads to extended funding for studies investigating an increase in efficiency and maximizing profit. Therefore, AI applications will be dominated by commercial interests, and SDG goals will not be considered.

Inequality. Vinuesa et al., 2020 found that several aspects of AI enhance disparities in society. First of all, issues relevant to researchers are often the ones close to where research institutions are located. For example, automated agriculture is a big topic in wealthy countries, whereas only a few studies focus on issues in developing countries with limited AI research. Additionally, AI changes job availability towards highly educated people. Countries with insufficient educational standards are limited by a knowledge gap, and AI development will stagnate. Similarly, it has been shown that salary gaps within nations between highly educated and low educated have increased in the last decades (4).
Alone the technical limit of computational power can act as a large inhibiter of the SDGs targets. The largest super-computing sites are located in the US, Europe, China, and Japan, giving researchers access to a fast computation of large data sets. Less wealthy countries are limited by computational power and even hardware, and inequality is, therefore, threatened to increase.

An evolving field in AI research is the ethical approach of AI (5). How can we define inequality? Should an algorithm be fair towards an individual or towards a group? How can we even tell an algorithm what fairness is? As mentioned above, inequality is often introduced due to inherent bias, and minorities are most vulnerable towards those. To address the SDG targets, the fairness of AI is an essential aspect of research.

Climate. As mentioned above, AI, big data, and high-performance computing have enabled scientists to better understand the climate system and to monitor possible threats towards a healthy environment. On the other hand, high demand in computational power and, therefore, energy has a negative impact on the climate action SDG. Furthermore, the usage of AI might lead to misusages to further exploit resources.

Society. One of the biggest challenges of AI is the acceptance of new technologies in society. An essential element to overcome prejudices is enough education for a basic understanding of the concepts. Notably, governments and companies’ credibility to not abuse collected data can reduce the success of AI in some countries. It is important to mention, though, that these concerns are very relevant and should not be discarded. The term “big nudging” has been introduced, for example, to describe the bias social media adds by showing specific newspaper articles or videos of newspapers or individuals with similar political interests as the user. This nudges the user towards a political opinion without the user’s active involvement.

AI is a dynamic and rapidly evolving field of science with an incredible number of applications in real life, many of which have already been implemented. In the phase of the digital revolution and the fast-changing circumstances, regulations on a national and international are too slow their implementations. Laws that both protect the public and prevent an increase in inequality but leave space for innovation are needed to ensure that the impact of AI on the SDGs remains mostly positive.

It remains to be seen, whether governments, companies, non-governmental organizations, and the public will use AI and big data to create a “better” world by addressing a majority of the SGDs. AI opens a lot of possibilities, but missing regulations and bias in training data may prevent a positive impact. Especially in the most vulnerable parts of our society.

1 Vinuesa, R., Azizpour, H., Leite, I., Balaam, M., Dignum, V., Domisch, S., … & Nerini, F. F. (2020). The role of artificial intelligence in achieving the Sustainable Development Goals. Nature Communications, 11(1), 1-10.

2 “AI is sending people to jail—and getting it wrong”. MIT Technology Review. Karen Hao, 2019.

3 Bolukbasi, T., Chang, K.-W., Zou, J., Saligrama, V. & Kalai, A. (2016). Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. Adv. Neural Inf. Process. Syst. 29, 4349–4357.

4 Brynjolfsson, E. & McAfee, A. (2014). The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies (W. W. Norton & Company).

5 “Microsoft Research Webinar: Machine Learning and Fairness”. Wallach and Wortmann Vaughan, 2019.