The most evident and straightforward area for ML implementation in ESG is ecology, which forms the first block of SDG targets in Figure 1. According to [15], AI methods may provide up to 93% of SDG targets for this domain. Examples include climate action [21], disaster monitoring [22], renewables [23], pollution reduction [24], forest restoration [25], and biodiversity preservation [26]. Many of these cases may be addressed by acquiring satellite or aerial images and subsequent procession using computer vision (CV) algorithms. Such images often include both optical and multispectral data and can be augmented with other remotely measured data [27]. Another major piece of information comes from tracking cameras (stationary or mounted on unmanned aerial vehicles (UAVs)) for wildlife observation of natural habitats [28]. Typically, CV algorithms rely on convolution neural networks, with the most classical architectures being the You Only Look Once (YOLO) family [29] and the Single-Shot MultiBox Detector (SSD) family [30].

A practical example of ML implementation in the environmental context is [31], in which the authors investigate the distribution of the plant Heracleum Sosnowskyi, which is considered an invasive species in northern and central Russia and some parts of Europe. Originally, the plant habitat was subalpine meadows and highland forests in the Caucasus region and northern parts of the Middle East, but in the 1970s, it was introduced in Europe as a silage plant and quickly spread in rural areas [32]. Apart from its aggressive competition with native plants, this hogweed contains phototoxic furanocoumarins [33], which greatly increases photosensitivity to ultraviolet light, and thus even a small amount of plant juice may cause severe sunburns when applied to human skin.

In [31], the authors proposed several prediction models based on publicly available environmental data to estimate whether hogweed will grow in certain areas by the year 2060. They implemented several climate models with the worst and best scenarios in terms of carbon emissions [34] and found that the most influential climatic parameters were the mean rainfall of the wettest months as well as the mean temperature of the wettest year quarter. The findings are alarming: There are significant chances of rapid plant propagation to the north of 60 latitude, where it does not currently grow (as of 2022) due to the cold climate [31]. Nevertheless, the spread of hogweed can be monitored and controlled by assessing the flora images obtained from UAVs. Such drones can automatically fly over large swaths of land, scan areas for the plant’s presence, and inform local forestry companies about possible actions (Figure 2).

Satellite images are most useful for simultaneously assessing huge portions of territory. Such images are captured with special sensors (mounted on Maxar, Landsat, Sentinel, WorldView, and similar spaceborne platforms) and can be obtained for almost any place on the planet with the required resolution. Additionally, the data can be captured for the same land repeatedly for a duration of several months or years, thus allowing for the observation of the area evolution both in the short term (wildfires, floods, etc.) and the long term (climate change, human activities, etc.). It is also beneficial that the basic satellite data (Landsat, Google Earth, etc.) is publicly available and thus can be readily used in achieving SDG targets.

For example, in forest management, ML algorithms are widely useful for the evaluation of the leaf area index, vegetation structure, moister level, and number of trees per acre [35,36,37,38]. These parameters provide a good health indication for a particular forest and may be used to evaluate the overall situation in the considered region [39]. Another closely related topic is the carbon neutrality target [40]. The background idea is that the emissions can be steadily compensated via vegetation CO₂ absorption. The overall area neutrality is evaluated in three steps. First, carbon dioxide spread is simulated using special biogeochemical models (Forest-DNDC) [41]. Second, plant taxation is performed for the area in question. Third, the capabilities of identified flora are compared with the presented levels of emissions (Figure 3). Therefore, one can easily assess whether more trees and bushes are required to compensate for the local CO₂ levels. When extrapolated on the scale of the whole country, the obtained data can be used for carbon reduction targets and provide the basis for domestic and international emission trading. A similar approach can also be used for other noncarbon sources of pollution [42].

Overall, the environmental domain is a good source of practical cases, many of which can be solved by using CV methods. The obtained information can provide either a basic insight into the object of study or be utilized as input data for prediction or classification problems. Since environmental awareness has become crucial in most societies, in the coming years, there will likely be many new cases suitable for the implementation of AI tools.

It is assumed that AI may be beneficial for 82% of social SDG targets (Figure 1), for example, SDG 1—eliminating poverty, SDG 4—quality education, and SDG 6—clean water and sanitation [15]. The recent SARS-CoV-2 pandemic also highlighted the necessity of disease spread modeling [43]. As can be seen from Figure 1, some society SDG targets are also tied to the environment sector, in particular, smart cities. Such cities are expected to be technologically advanced in gathering digital data and efficiently using them to govern urban infrastructure and assets [44]. Some relevant AI-supported cases are given below:

• Smart electric grids can be adjusted to customer demand and switch between various generation sources (wind farms, solar panels, thermal energy, etc.) [45,46]. A similar approach can be adopted to manage other resources like water or heating for intelligent buildings [47];
• Predictive maintenance of crucial infrastructure prevents accidents and disruptions [48];
• A human-centered environment leads to transparent residents–authorities interaction [49];
• Advanced healthcare helps in disease diagnostics and treatment planning [50,51];
• Optimized logistics and scheduling improves transportation sustainability [52];
• Traffic lights adjust to the road situation and prevent traffic jams [47].

The above cases are mainly solved by using various optimization and classifying/clustering algorithms (e.g., random forest [53], boosting [54], deep learning neural networks [55], and others [56]). CV methods are also extensively employed for smart city problems due to the wide use of cameras in modern urban environments [57].

Land management and cadaster revision also benefit from AI. First, satellites provide data images of the urban and rural areas in question. Next, various land objects like crop fields, townhouses, parks, or industrial plants are identified and classified in photos. Lastly, cadaster data are compared with the actual information on the object, and then the register is updated if necessary [58]. The advanced algorithms can extend even further and provide landowners with suggestions on different variants for subsequent area development, thus improving asset handling (Figure 4). For instance, when evaluating plans for building an apartment block, a developer can use ML tools to find and check similar edifices to verify the construction estimate and find the best construction team or building material vendor.

Another case is the use of AI for naval logistics in ice-covered seas. As global warming makes the Arctic region more suitable for shipping, this route becomes an attractive alternative for moving cargo between Europe and Asia. In particular, the distance between Murmansk (Russia) and Yokohama (Japan) is almost 7000 nautical miles shorter along the Northern Sea Route than the traditional path through the Suez Canal. Thus far, several works studied sea-ice behavior by combining classical ocean models and ML methods [59,60]. The main challenge here is to provide a computationally lightweight and reliable forecast with the scarcely available weather data so that the ship crew will be able to set the best course through the moving ice floes. For example, Ref. [61] reports such a cost-effective model based on U-Net architecture, which operates in two different regimes and is capable of predicting the ice thickness for the next 10 days. In [61], the model was successfully tested for the Barents Sea, the Labrador Sea, and the Laptev Sea regions, and its practical performance for ocean navigation was assessed.

On the whole, the most popular objects of study in the social domain are different aspects of smart cities’ performance. Similar to environmental tasks, CV algorithms can be used here; however, the most practical cases are associated with optimization, prediction, and classification/clustering problems. As the urban population steadily grows, and IT technologies become more common, it should be expected that more practical AI cases will also arise with time.

In order to verify the obtained results, the researchers determined Amundi Asset Management ESG scores [66] (which are routinely found by manually processing the company information) and compared them with the esgNLP calculations. While the results were different for the investigated companies, the study showed that the organizations with higher esgNLP scores on average had higher Amundi ESG scores as well [63].

A similar approach was applied in [2], where the authors focused on a comparison of AI-based scores from Truvalue Labs [67] and “traditional” assessment scores from MSCI ESG [68]. In contrast to the previous work, the results show a weak correlation between scores, which was attributed to a potentially more biased approach in a traditional assessment (larger companies can provide more data for reports used in assessment). The other possible reason was the information source. In the experiment, AI-based scores were mainly calculated from publicly available data, while the traditional assessment heavily relied on company disclosure [2]. It should be noted that, despite contradictions in scores, ML methods still have a strong potential in ESG assessment. Most likely, further development of NLP models will make them a valuable tool for investors and asset managers by providing supplementary data to make better decisions.

The internal aspect of governance, which is related to company self-assessment [69], focuses on the efficiency of its policy and bureaucracy. Once again, the most valuable tools here are based on NLP technology, which automates routine tasks such as checking document consistency [70], generating reports according to preset templates [71], implementing search engines [72], establishing knowledge bases [73], and supporting chatbots to assist employees [74]. However, NLP methods should be distinguished from standard robotic process automation (RPA) [75], which also simplifies everyday tasks. The main difference is that RPA simply remembers the order of user actions (such as keystrokes or mouse movement) and repeats them again and again. Thus, it is effective when one has to download attachments from 1000 emails but fails if it is required to sort these files according to some nontrivial rule.

The main AI technology for the governance domain is NLP, which focuses on text processing. While the external aspect of the governance domain still requires significant research, internal applications have already proven their efficiency. With the recent breakthroughs in heavyweight language models, a new era of practical AI cases is already starting to emerge.

The other issue is ethics-based AI auditing, which needs to be addressed if ML algorithms are expected to make fair and safe decisions unbiased with unrepresentative datasets used for their training [15,76]. In [5], at least 173 AI auditing frameworks were mentioned, but the lack of universally adopted metrics and standards was highlighted. Overall, it is agreed that several criteria should be applied to AI systems such as nonmaleficence, transparency, responsibility, fairness, and privacy [77]. The designed algorithms should also be robust to cope with possible discrepancies in the input data, especially when operating on some highly important infrastructure [78]. The main challenge here is to transform the general framework criteria into practical measures with strict wording and quantitative description so that the stakeholders could apply them on a day-to-day basis. The most reasonable way would be to select one of the frameworks and adopt it as legislation with the help of an AI expert society.

There is also a large discussion on the social impact of AI as a disruptive technology [79]. Despite creating new IT jobs, ML typically leads to the automatization of certain business and technological processes. As a result, companies consider closing low-skill positions, and some of their employees suffer pay cuts or even lose their jobs, which became obsolete [15]. In addition to that, the companies failing to timely implement AI-enhanced production potentially become less competitive in the market and have fewer chances to obtain funding from investors. It should also be noted that private enterprises primarily focus only on the areas that provide clear economic effects (for example, production optimization in an assembly plant). At the same time, other problems with less evident impact on the business model become less attractive (e.g., a smart office for an oil rig), which leads to nonuniform progress toward some SDG targets [80]. This is also worsened by varying funding capabilities in that large enterprises have more opportunities to invest in AI than small businesses [15]. Likely, the best approaches here are improving public awareness of AI, promoting employees to develop professional skills in ML, and increasing funding of not-for-profit applications.

Many ML implementations such as tracking and person recognition can be seen as ethically questionable, especially when combined with calculating citizen scores and rankings [81]. Since different countries have varying cultural and political backgrounds, an application of public-related ML algorithms (e.g., recommendations on a social network) that worked fine in one environment may lead to dramatic consequences in other cases [15]. So, such tools should be used with great care and tested in advance with representatives of the new audiences.

Another interesting case is linked with fake news, which is defined as presenting misleading or false information that is claimed to be true. Such news appears on a routine basis in media often spreading around hot topics like celebrity scandals, rigged elections, or proposed tax cuts [82]. Once a rumor is out, it is very hard to deal with, as the public is prone to sensations and keeps circulating false data. The situation has become even worse with AI advancement since this technology provides quite powerful tools for fake news generation (especially with generative adversarial networks (GANs)) [83]. GANs are capable of efficiently mimicking training material to look almost authentic to the observer. GANs can easily adjust video and images by replacing faces and backgrounds (Deepfake), thus causing potentially huge reputation or economic harm [84]. Thus far, several practical methods have been proposed to detect and deal with fake graphic content [85,86], but some fundamental research is also required for better performance [87,88]. Similar fact-checking tools are available for text content as well [89,90].

The other question that requires a separate discussion is ChatGPT [91]. This AI-powered tool effectively mimics an interlocutor and is capable of keeping a meaningful discussion, writing lyrics, or generating a basic programming code. Undoubtedly, ChatGPT has tremendous potential; however, the surrounding hype and public expectations hinder an assessment of its practical value for the industry right now (as of the beginning of 2023). In order to work properly in the host company environment, ChatGPT should be trained and have constant access to the company’s internal data and documentation. However, such access would be strongly opposed by the company’s information security service (ISS), whose primal goal is to prevent the leakage of any sensitive information such as personal data, financial performance, or procedural knowledge (“knowing-how”). Thus, the first step here should be to adjust the expectations of both potential ChatGPT users and ISS. Based on that, the company’s IT landscape should be redesigned so that ChatGPT would always only assess harmless general data but still provide great practical experience to users by, for example, helping to manage meetings and write memos.

Lastly, it should be noted that ML methods have become even more energy-demanding. For example, training a state-of-the-art GPT-3 NLP network requires almost 936 MWh—the same amount of power as 468,000 average electric kettles (2 kWh) or 12,480 lithium-ion Tesla car batteries (75 kWh). It is estimated that, by 2030, information and communications technologies (including AI) would need 20% of the world’s electricity, which would lead to a great carbon footprint unless this issue is addressed in advance [92]. The computational difficulty of many ML-related problems also drives a demand for more graphical processing units, typically used in calculations. Their production involves the utilization of toxic materials and thus has a strong negative impact on the environment [93]. In particular, arsenic, phosphine, sulfidic and hydrofluoric acids are widely used in various steps of the semiconductor fabrication process. These materials pose a direct risk to human health and are also reported to increase the chances of cancer later in life. The above drawbacks can be partially solved by the implementation of the so-called “green AI”, which uses specially designed neural networks with lower power consumption and, consequently, lesser ecological impact [94].