Resilient and Sustainable AI. Positioning paper on the relation of AI, resilience and sustainability

Artificial intelligence, resilience, and sustainability are major concepts of the current debate on the future of work and life. In the industrial landscape, the convergence of AI, sustainability, and resilience is ushering in transformative changes that will reshape traditional paradigms. Each of these elements itself, and to an even greater extent together, are emerging as key drivers for the next era of industrial growth. Some well-known examples of the pivotal role of AI in Industry are:

Even though everybody talks or writes about AI, most people have no concrete idea of what the term actually stands for. To talk about artificial intelligence, it is therefore necessary to define it in the first place. In the current period, with new software products based on machine learning (ML) technology appearing almost daily, ML dominates the public’s perception and broad understanding of what AI is.

In a JRX technical report by the EU [16], various definitions were taken up and analysed by AI Watch with the goal of sharpening the vague concept of AI. It concludes in stating, that AI is a catch-all term and can mean different things:

“AI is a generic term that refers to any machine or algorithm that is capable of

This definition approach essentially builds on the idea of a description of the capabilities of the technology. This means that the AI term can thus encompass any technology that is capable of observing, learning, and, based on built-up knowledge, making decisions, and recommending or implementing them. Hence, we can see that multiple technologies fall under such a broad AI definition.

AI Watch develops an AI taxonomy based on this definition to cluster various AI approaches. Core areas of AI (Reasoning, Planning, Learning, Communication, Perception) are distinguished from transversal areas (Integration & Interaction, Services, Ethics & Philosophy) and sub-areas are named, such as knowledge representation in the area of reasoning [16]. Beside the more or less academic discourse about the question “What is AI?”, AI is getting to practice, which results in a political as well as legal definition of AI.

In the course of developing rules for AI (“LAYING DOWN HARMONISED RULES ON ARTIFICIAL INTELLIGENCE (ARTIFICIAL INTELLIGENCE ACT) AND AMENDING CERTAIN UNION LEGISLATIVE ACTS”) by the EU, three technologies are fundamentally distinguished:

“(a) Machine learning approaches, including supervised, unsupervised and reinforcement learning, using a wide variety of methods including deep learning;

(b) Logic- and knowledge-based approaches, including knowledge representation, inductive (logic) programming, knowledge bases, inference and deductive engines, (symbolic) reasoning and expert systems;

(c) Statistical approaches, Bayesian estimation, search and optimization methods.” [15]

Both presented definitions are highly overlapping and demonstrate somehow the close relation between legal/political interpretation and a specific academic debate. Again, this paper does not aim to present a full history of AI or a literature-based analysis of the current AI definition debate. From the authors point of view, the presented definitions are a practical ground for current developments of AI (Europe) in research and industry because of its role as legal framework. It sets the basics to understand that the term “AI” is generic, but refers to specific properties of technical systems, which can manifest completely differently – as a robot or as a statistical model. The expressions and intensity of the abilities are not further defined here. It describes what a technical system does (observe, learn, propose or implement decisions based on collected knowledge) to determine if this system is artificially intelligent.

Resilience is a term that is used in many contexts, e. g. psychology. It probably originally stems from the field of physics [25]. In the following, we will refer to economic resilience, which, likewise, can have many different forms. Resilience occurs within companies, as well as within value-added networks, and entire economies. The term was first used by Holling; stating that economic resilience “refers to the extent to which a shock can be absorbed by a local stable domain before it is induced into some other stable equilibrium” (Modica & Reggiani, 2015). Looking at the term through a systemic lens (e.g., in addition to 6), resilience is the ability of a system,

Similar to the AI definitions, the given approach to define resilience bases on an ability of a system (in this case, not necessarily a technical system). Unlike attributing the what a system must do (see AI), this definition rather focuses on the consequences or the result. This definition stands in contrast to the definition of organizational resilience in standardization. E.g., in the ISO standards’ definition, resilience isn’t an element of corporate control but emerges from the application of these control elements. [19] The question is whether a result of applying unspecified mechanisms can be a capability of systems? The aforementioned definition suggests precisely this – but above all, it suggests that systems can exhibit mechanisms that produce resilience. To address this challenge, resilience can be defined as responsiveness:

“Resilience is the responsiveness of the system, i.e., its elasticity or capacity to rebound after a shock, indicated by the degree of flexibility, persistence of key functions, or ability to transform […]” [27]

The mechanisms to achieve this can be termed as a resilience strategy. This, essentially, describes the deliberate approach to achieving resilience and can be clustered into three areas:

If a system possesses mechanisms that serve this purpose and are functional, then a system can be described as resilient. Such mechanisms are, for instance, part of supply chain research, which, for example, views “Situation Awareness” as a metric for “Readiness” as preparation see [19]. On the other hand, there are a number of standards that deal with resilience strategies. For companies, there’s ISO 22316, which suggests what to do or exactly how to achieve resilience, e.g., that the management should establish relevant business resilience goals and formulate assessment criteria for evaluating resilience attributes [21]

In addition to these standards, there are several others, closely tied to this capability, such as Compliance Management System (CMS) according to ISO 37301, Risk Management: ISO 31000, Knowledge Management: DIN EN ISO/IEC 27.001:2017; ISO 30301:2019; DIN ISO 30401:2021, Supply Chains: ISO 28000; DIN ISO 20400:2021 ISO/TS 22318. As already explained in the introductory words, this presentation is not comprehensive but provides a brief insight without even approaching the state of research.

After AI and resilience, sustainability remains to be addressed, an equally complex term. From the authors perspective, the concept of sustainability only becomes meaningful when it is understood as a sustainable development, meaning a process – and this is also how some other authors perceive it. Sustainability itself is a “paradigm for thinking about the future in which environmental, societal, and economic considerations are balanced in the pursuit of improved quality of life. The ideals and principles behind it rest on broad concepts such as intergenerational equity, gender equity, social tolerance, poverty alleviation, environmental preservation and restoration, natural resource conservation, and building just and peaceful societies” [23].

In the context of the so-called Triple Bottom Line and from a business perspective, sustainability is often understood as follows:

However, the “inventor” of the TBL, John Elkington, clarified in 2018 that, “[…] the original idea was broader still, urging businesses to track and manage economic (not just financial), social, and environmental value added – or destroyed.” [14] Thus, sustainability can’t just be measured by profit and loss, but by the well-being of people, now and in the future. The TBL emphasized radical transformation, market disruption, asymmetric growth (wherein non-sustainable sectors are deliberately marginalized), and the amplification of emergent market strategies (Elkington, 2018). Moreover, the TBL can be understood as a hierarchical system, not an equal one, because: Ecological conditions and resources serve as foundational elements for both life and societal structures. A well-operating society is pivotal for addressing social necessities and instituting an economic framework. This economic structure facilitates both resource utilization and prosperity, catering to material requirements [20].

From the political debate on sustainability, the concept of sustainable development marks a milestone. Sustainable development is a process filled with life in the context of the UN, especially the Brundtland Commission, which states that sustainable development addresses contemporary requirements while ensuring that future generations retain the capacity to fulfill their respective needs [5]. Without delivering a complete historical account (there’s plenty of literature for that, e.g., [18]), these developments culminated in the Sustainable Development Goals, the UN’s 17 sustainability targets. These are:

Combining these objectives with the TBL, goals 6, 13, 14, 15 (environment) can be seen as the foundation for goals 1, 2, 3, 4, 5, 7, 11 and 16 (society), which in turn enable goals 8, 9, 10, 12 (economy) and can be achieved through 17 (cooperation). As an example gender equality (#5) can be seen as a driver for high quality education (#4), enabling decent work and economic (#8) in the long term [24]. However, a too simplistic understanding would neglect the high interdependence of individual factors, as cause and effect relations are often precisely the opposite, e.g., sustainable production allows life on land and so on.

Having introduced the third term, the next sections will relate two terms to each other, finally focusing on the relationship between AI, sustainability, and resilience and discussing the role of AI.

Resilience and sustainability seem somewhat contradictory in the presented version. While sustainable development takes a very long-term view, resilience is more focused on narrow time horizons. This is especially the case when we correctly consider the type of response. Resilience is a reactive ability to a very abrupt disruption or crisis, unlike sustainable development which focuses on changes observed over very long processes. However, both concepts consider the maintenance of systems and their performance, but the time horizons they consider are very different.

Nevertheless, both concepts are related. A sustainable system should be able to handle shocks and crises. “In political interpretation, resilience and sustainability are two sides of the same coin. Only sustainable value chains remain robust and intact in the long run. It is therefore essential that companies keep an eye on the impact on people, the environment, and the climate in their activities” (4, translated). The logical relationship that arises from this is: “Resilience represents a necessary, but not sufficient, condition for sustainability.” (7, translated). The logical relationship that arises from this is: “Resilience represents a necessary, but not sufficient, condition for sustainability” (Translated from 7).

In short: Not every resilient system is sustainable, but every sustainable system is resilient. This dependency and interdependence can lead to elements of sustainable development, being part of resilience strategies as they can be found in the BMWK-Whitepaper in context of Platform Industry 4.0 [3]. Sustainability is one of the levels of the resilience strategy, which spans the various impact phases of resilience. The social dimension of sustainability – health protection, societal benefit, working conditions – is particularly emphasized. In other words, in Industry 4.0, social sustainability constitutes a sufficient condition for resilience

Resilience and AI can be related in various ways. As already described, AI always pertains to a technical system’s ability to learn/act, while resilience addresses a systems responsiveness. From this, three potential connections arise:

Regarding the point that AI is used for resilience, there are numerous (German or EU) research projects and publications (e.g. [2, 10, 31]). Projects and their areas can be taken from the following table (Tab. 1):

AI can be used in all phases of the resilience strategy. The most intuitive applications involve AI-supported prediction systems for specific scenarios. This is particularly useful when risks for a shock can be foresight – hence, enough data with enough information is available to make such predictions.

Besides AI’s beneficial power to strengthen resilience, AI system may also have other roles. The second role of AI is, that AI is realized as a trigger for shocks on industries or companies. A publicly discussed example of such kind of shock, was the discussion about ChatGPT and its impact on stocks in the education market, which (temporarily) plunged². This example is a prototype for public discourses about the impact of AI, which moves between from euphoria to dystopia (in this case near economic dystopia). Reality probably lies somewhere in the middle, as AI development influences society in general – just as society and economic systems influence technological developments, see also [33]. The interdependencies are complex, making it not unlikely that AI developments will cause shocks to hit companies and sectors.

Hence, it seems logical to examine how resilience can be built up for such case. Initially, understanding what AI is and that it will induce change should help organizations prepare for potential AI-driven shocks. Building this competence or accessing educational resources (also pertaining to social sustainability), is useful throughout all resilience phases. The level of competency being built also aids during the adaptation phase, leveraging the shock for potential adaptations and restructuring measures.

Lastly, the resilience of AI systems themselves can be examined. Why? Since the beginning of computing, various significant errors have happened due to numerical inaccuracies. A particularly infamous incident was the self-destruction of an Ariane 5 rocket in 1996. Such events have driven the establishment of a dedicated research area that concentrates on creating and refining resilient algorithms, a niche within the realm of numerics. This needs to be extended on AI Systems to prevent the given risks [30].

As AI systems increasingly permeate our daily lives and work, it becomes essential to scrutinize their resilience to prevent potential harm to social systems or the organizations we work in. Wittenbrink et al. [30] tackled this challenge, concluding that five factors define resilient AI:

These points refer, among other things, to the data foundation upon which any AI is based, its origin, and conformity (Governance Risk Compliance or GDPR), IT security, and the model’s explainability and resilience, or that of the technical system. Research around this topic is, to the authors knowledge, still in its infancy, and the question of extending and specifying these points is probably an ongoing process. In the future, questions about model transparency and the focus on the common good will likely arise. I will revisit the question of AI system resilience when introducing the connection between AI and sustainability, as resilience and sustainability are logically connected, and some readers might have noticed certain overlaps.

There are several works about sustainability and AI. As an example, the definition of sustainable AI is still being formulated. Bjørlo et al. and Halsband focuses on the idea that sustainable AI should meet current needs without compromising the future – in addition to the presented definition of sustainability. Some researchers have identified boundary conditions for sustainable AI, including diversity, trust, and the capacity for self-organization and learning.

One interesting work is “Sustainability Criteria for Artificial Intelligence” („Nachhaltigkeitskriterien für künstliche Intelligenz“) authored by Friederike Rohde, Josephin Wagner, Philipp Reinhard, Ulrich Petschow, Andreas Meyer, Marcus Voß, and Anne Mollen [26]. The paper delves into the realm of artificial intelligence (AI), particularly machine learning (ML) and sustainability. The increasing deployment of AI systems has sparked debates globally about their societal, environmental, and economic impacts. Concerns include non-transparent decision-making processes, discrimination, rising energy consumption, greenhouse gas emissions during AI model development, and broader consequences on labor markets, consumption patterns, and the market power of large corporations. The authors aim to provide a perspective on sustainable AI, connecting it with discussions on AI ethics, Green AI, and the sustainability of AI and advocate for an embedded perspective on sustainable AI, viewing technology as socially shaped and shapable, not just a neutral tool [26].

From a sustainability perspective, AI systems present multiple challenges. One of these significant challenges is the interdependency between different SDGs and AI. Direct and indirect impacts of AI on SDGs need to be considered³. Vinuesa et al. [28] note that AI could aid in achieving 79% of the Sustainable Development Goals, like alleviating poverty and improving education. However, it could also hinder 35% of these goals by, for instance, consuming vast natural resources, propagating biases against gender equality, or bolstering autocratic regimes. While AI in sectors like smart manufacturing can boost productivity and conserve resources, it might lead to job losses in areas like finance⁴. However, the work of Rohde et al. (2021) includes a systematic overview of impacts throughout the AI lifecycle. Further Rohde et al. presents conceptual ideas for a comprehensive sustainability assessment of AI. It introduces thirteen sustainability criteria and five cross-sectional criteria with corresponding indicators based on existing scientific and societal discourses on the impacts of AI [26]:

In the previous chapters, numoerous definitions and explanation approaches for the understanding of the major concepts of AI, sustainability and resilience and their interconnections were given. To sum it up, we can state that AI is a generic term for technical systems with special capabilities (observe, learn, act), Resilience refers to a responsiveness of a system, and sustainability or sustainable development addresses contemporary requirements while ensuring that future generations retain the capacity to fulfill their respective needs.

AI, understood as a technical system, can be resilient, and in an extended sense, sustainable (with sustainability implying resilience, as demonstrated). Hence, AI systems need to ensure to fulfil or not to violate contemporary requirements on sustainability (given above). Therefore, resilient AI systems are a necessary but not sufficient condition for sustainable AI systems. Technical systems are integral components of socio-technical systems, suggesting that these technical systems impact and are interdependent with social, ecological, and economic systems throughout the AI systems' lifecycle. The sustainability index approach introduced for AI [26] is a pivotal step in identifying and measuring these interdependencies. To integrate the aforementioned works, the author proposes that the sustainability and inherent resilience of AI systems, as well as their impact concerning sustainability factors, are sufficient criteria for evaluating AI’s overall sustainability. From this perspective, it’s unsurprising that many resilient factors are already evident within the presented sustainable AI index, such as the data foundation in AI systems documentation and Technical Reliability. This suggests that some resilience factors might need more in-depth consideration in terms of [30]:

Moreover, we must recognize that certain AI technologies could be leap innovations (market shocks) and might be unpredictable. We could also face challenges when trying to identify rebound effects. AI technologies will invariably introduce potential shocks to systems. To prepare for these shocks, developing competencies has been showcased as an effective way to navigate such potential AI-induced disruptions (and perhaps even broader ones), thus fostering sustainable socio-technical systems. As a result, the index should also give significantly more weight to: Competency development for co-creating AI systems as a pertinent factor.

Besides those factors, more generally, we may need to considerate and classify the factors in different level. Finally, sustainable AI systems needs to be resilient – it is a logical consequence.