This is an excerpt from my latest contribution on responsible artificial intelligence (AI).
Suggested citation: Camilleri, M. A. (2023). Artificial intelligence governance: Ethical considerations and implications for socialresponsibility. Expert Systems, e13406. https://doi.org/10.1111/exsy.13406
The term “artificial intelligence governance” or “AI governance” integrates the notions of “AI” and “corporate governance”. AI governance is based on formal rules (including legislative acts and binding regulations) as well as on voluntary principles that are intended to guide practitioners in their research, development and maintenance of AI systems (Butcher & Beridze, 2019; Gonzalez et al., 2020). Essentially, it represents a regulatory framework that can support AI practitioners in their strategy formulation and in day-to-day operations (Erdélyi & Goldsmith, 2022; Mullins et al., 2021; Schneider et al., 2022). The rationale behind responsible AI governance is to ensure that automated systems including ML/DL technologies, are supporting individuals and organizations in achieving their long terms objectives, whist safeguarding the interests of all stakeholders (Corea et al., 2023; Hickok et al., 2022).
AI governance requires that the organizational leaders comply with relevant legislation, hard laws and regulations (Mäntymäki et al., 2022). Moreover, they are expected to follow ethical norms, values and standards (Koniakou, 2023). Practitioners ought to be trustworthy, diligent and accountable in how they handle their intellectual capital and other resources including their information technologies, finances as well as members of staff, in order to overcome challenges, minimize uncertainties, risks and any negative repercussions (E.g. decreased human oversight in decision making, among others) (Agbese et al., 2023; Smuha, 2019).
Procedural governance mechanisms ought to be in place to ensure that AI technologies and ML/DL models are operating in a responsible manner. Figure 1 features some of the key elements that are required for the responsible governance of artificial intelligence. The following principles are aimed to provide guidelines for the modus operandi of AI practitioners (including ML/DL developers).
Figure 1. A Responsible Artificial Intelligence Governance Framework
Accountability and transparency
“Accountability” refers to the stakeholders’ expectations about the proper functioning of AI systems, in all stages, including in the design, creation, testing or deployment, in accordance with relevant regulatory frameworks. It is imperative that AI developers are held accountable for the smooth operation of AI systems throughout their lifecycle (Raji et al., 2020). Stakeholders expect them to be accountable by keeping a track record of their AI development processes (Mäntymäki et al., 2022).
The transparency notion refers to the extent to which end-users could be in a position to understand how AI systems work (Andrada et al., 2020; Hollanek, 2020). AI transparency is associated with the degree of comprehension about algorithmic models in terms of “simulatability” (an understanding of AI functioning), “decomposability” (related to how individual components work), and algorithmic transparency (this is associated to the algorithms’ visibility).
In reality, it is difficult to understand how AI systems, including deep learning models and their neural networks are learning (as they acquire, process and store data) during training phases. They are often considered as black box models. It may prove hard to algorithmically translate derived concepts into human-understandable terms, even though developers may use certain jargon to explain their models’ attributes and features. Many legislators are striving in their endeavors to pressurize AI actors to describe the algorithms they use in automated decision-making, yet the publication of algorithms is useless if outsiders cannot access the data of the AI model.
Explainability and interpretability
Explainability is the concept that sheds light on how AI models work, in a way that is comprehensible to a human being. Arguably, the explainabilty of AI systems could improve their transparency, trustworthiness and accountability. At the same time, it can reduce bias and unfairness. The explainability of artificial intelligence systems could clarify how they reached their decisions (Arya et al., 2019; Keller & Drake, 2021). For instance, AI could explain how and why autonomous cars decide to stop or to slow down when there are pedestrians or other vehicles in front of them.
Explainable AI systems might improve consumer trust and may enable engineers to develop other AI models, as they are in a position to track provenance of every process, to ensure reproducibility, and to enable checks and balances (Schneider et al., 2022). Similarly, interpretability refers to the level of accuracy of machine learning programs in terms of linking the causes to the effects (John-Mathews, 2022).
Fairness and inclusiveness
The responsible AI’s fairness dimension refers to the practitioners’ attempts to correct algorithmic biases that may possibly (voluntarily or involuntarily) be included in their automation processes (Bellamy et al., 2019; Mäntymäki, et al., 2022). AI systems can be affected by their developers’ biases that could include preferences or antipathies toward specific demographic variables like genders, age groups and ethnicities, among others (Madaio et al., 2020). Currently, there is no universal definition on AI fairness.
However, recently many multinational corporations have developed instruments that are intended to detect bias and to reduce it as much as possible (John-Mathews et al., 2022). In many cases, AI systems are learning from the data that is fed to them. If the data are skewed and/or if they comprise implicit bias into them, they may result in inappropriate outputs.
Fair AI systems rely on unbiased data (Wu et al., 2020). For this reason, many companies including Facebook, Google, IBM and Microsoft, among others are striving in their endeavors to involve members of staff hailing from diverse backgrounds. These technology conglomerates are trying to become as inclusive and as culturally aware as possible in order to minimize bias from affecting their AI processes. Previous research reported that AI’s bias may result in inequality, discrimination and in the loss of jobs (Butcher & Beridze, 2019).
Privacy and safety for consumers
Consumers are increasingly concerned about the privacy of their data. They have a right to control who has access to their personal information. The data that is collected or used by third parties, without the authorization or voluntary consent of individuals, would result in the violations of their privacy (Zhu et al., 2020; Wu et al., 2022).
AI-enabled products, including dialogue systems like chatbots and virtual assistants, as well as digital assistants (e.g. like Siri, Alexa or Cortana), and/or wearable technologies such as smart watches and sensorial smart socks, among others, are increasingly capturing and storing large quantities of consumer information. The benefits that are delivering these interactive technologies may be offset by a number of challenges. The technology businesses who developed these products are responsible to protect their consumers’ personal data (Rodríguez-Barroso et al., 2020). Their devices are capable of holding a wide variety of information on their users. They are continuously gathering textual, visual, audio, verbal, and other sensory data from consumers. In many cases, the customers are not aware that they are sharing personal information to them.
For example, facial recognition technologies are increasingly being used in different contexts. They may be used by individuals to access websites and social media, in a secure manner and to even authorize their payments through banking and financial services applications. Employers may rely on such systems to track and monitor their employees’ attendance. Marketers can utilize such technologies to target digital advertisements to specific customers. Police and security departments may use them for their surveillance systems and to investigate criminal cases. The adoption of these technologies has often raised concerns about privacy and security issues. According to several data privacy laws that have been enacted in different jurisdictions, organizations are bound to inform users that they are gathering and storing their biometric data. The businesses that employ such technologies are not authorized to use their consumers’ data without their consent.
Companies are expected to communicate about their data privacy policies with their target audiences (Wong, 2020). They have to reassure consumers that the consented data they collect from them is protected and are bound to inform them that they may use their information to improve their customized services to them. The technology giants can reward their consumers to share sensitive information. They could offer them improved personalized services among other incentives, in return for their data. In addition, consumers may be allowed to access their own information and could be provided with more control (or other reasonable options) on how to manage their personal details.
The security and robustness of AI systems
AI algorithms are vulnerable to cyberattacks by malicious actors. Therefore, it is in the interest of AI developers to secure their automated systems and to ensure that they are robust enough against any risks and attempts to hack them (Gehr et al., 2018; Li et al., 2020).
The accessibility to AI models ought to be continuously monitored at all times during their development and deployment (Bertino et al., 2021). There may be instances when AI models could encounter incidental adversities, leading to the corruption of data. Alternatively, they might encounter intentional adversities when they experience sabotage from hackers. In both cases, the AI model will be compromised and can result in system malfunctions (Papagiannidis et al., 2023).
AI models have to prevent such contingent issues from happening. Their developers’ responsibilities are to improve the robustness of their automated systems, and to make them as secure of possible, to reduce the chances of threats, including by inadvertent irregularities, information leakages, as well as by privacy violations like data breaches, contamination and poisoning by malicious actors (Agbese et al., 2023; Hamon et al., 2020).
AI developers should have preventive policies and measures related to the monitoring and control of their data. They ought to invest in security technologies including authentication and/or access systems with encryption software as well as firewalls for their protection against cyberattacks. Routine testing can increase data protection, improve security levels and minimize the risks of incidents.
Conclusions
This review indicates that more academics as well as practitioners, are increasingly devoting their attention to AI as they elaborate about its potential uses, as well as on its opportunities and threats. It reported that its proponents are raising awareness on the benefits of AI systems for individuals as well as for organizations. At the same time, it suggests that a number of scholars and other stakeholders including policy makers, are raising their concerns about its possible perils (e.g. Berente et al., 2021; Gonzalez et al., 2020; Zhang & Lu, 2021).
Many researchers identified some of the risks of AI (Li et al., 2021; Magas & Kiritsis, 2022). In many cases, they warned that AI could disseminate misinformation, foster prejudice, bias and discrimination, raise privacy concerns, and could lead to the loss of jobs (Butcher & Beridze, 2019). A few commentators argue about the “singularity” or the moment where machine learning technologies could even surpass human intelligence (Huang & Rust, 2022). They predict that a critical shift could occur if humans are no longer in a position to control AI anymore.
In this light, this article sought to explore the governance of AI. It sheds light on substantive regulations, as well as on reflexive principles and guidelines, that are intended at practitioners who are researching, testing, developing and implementing AI models. It clearly explains how institutions, non-governmental organizations and technology conglomerates are introducing protocols (including self-regulations) to prevent contingencies from even happening due to inappropriate AI governance.
Debatably, the voluntary or involuntary mishandling of automated systems can expose practitioners to operational disruptions and to significant risks including to their corporate image and reputation (Watts & Adriano, 2021). The nature of AI requires practitioners to develop guardrails to ensure that their algorithms work as they should (Bauer, 2022). It is imperative that businesses comply with relevant legislations and to follow ethical practices (Buhmann & Fieseler, 2023). Ultimately, it is in their interest to operate their company in a responsible manner, and to implement AI governance procedures. This way they can minimize unnecessary risks and safeguard the well-being of all stakeholders.
This contribution has addressed its underlying research objectives. Firstly, it raised awareness on AI governance frameworks that were developed by policy makers and other organizations, including by the businesses themselves. Secondly, it scrutinized the extant academic literature focused on AI governance and on the intersection of AI and CSR. Thirdly, it discussed about essential elements for the promotion of socially responsible behaviors and ethical dispositions of AI developers. In conclusion it put forward an AI governance conceptual model for practitioners.
This research made reference to regulatory instruments that are intended to govern AI expert systems. It reported that, at the moment there are a few jurisdictions that have formalized their AI policies and governance frameworks. Hence, this article urges laggard governments to plan, organize, design and implement regulatory instruments that ensure that individuals and entities are safe when they utilize AI systems for personal benefit, educational and/or for commercial purposes.
Arguably, one has to bear in mind that, in many cases, policy makers have to face a “pacing problem” as the proliferation of innovation is much quicker than legislation. As a result, governments tend to be reactive in the implementation of regulatory interventions relating to innovations. They may be unwilling to hold back the development of disruptive technologies from their societies. Notwithstanding, they may face criticism by a wide array of stakeholders in this regard, as they may have conflicting objectives and expectations.
The governments’ policy is to regulate business and industry to establish technical, safety and quality standards as well as to monitor their compliance. Yet, they may consider introducing different forms of regulation other than the traditional “command and control” mechanisms. They may opt for performance-based and/or market-based incentive approaches, co-regulation and self-regulation schemes, among others (Hepburn, 2009), in order to foster technological innovations.
This research has shown that a number of technology giants, including IBM and Microsoft, among others, are anticipating the regulatory interventions of different governments where they operate their businesses. It reported that they are communicating about their responsible AI governance initiatives as they share information on their policies and practices that are meant to certify, explain and audit their AI developments. Evidently, these companies, among others, are voluntarily self-regulating themselves as they promote accountability, fairness, privacy and robust AI systems. These two organizations, in particular, are raising awareness about their AI governance frameworks to increase their CSR credentials with stakeholders.
Likewise, AI developers who work for other businesses, are expected to forge relationships with external stakeholders including with policy makers as well as with actors including individuals and organizations who share similar interests in AI. Innovative clusters and network developments may result in better AI systems and can also decrease the chances of possible risks. Indeed, practitioners can be in better position if they cooperate with stakeholders for the development of trustworthy AI and if they increase their human capacity to improve the quality of their intellectual properties (Camilleri et al., 2023). This way, they can enhance their competitiveness and growth prospects (Troise & Camilleri, 2021). Arguably, it is in their interest to continuously engage with internal stakeholders (and employees), and to educate them about AI governance dimensions, that are intended to promote accountable, transparent, explainable interpretable reproducible, fair, inclusive and secure AI solutions. Hence, they could maximize AI benefits, minimize their risks as well as associated costs.
Future research directions
Academic colleagues are invited to raise more awareness on AI governance mechanisms as well as on verification and monitoring instruments. They can investigate what, how, when and where protocols could be used to protect and safeguard individuals and entities from possible risks and dangers of AI.
The “what” question involves the identification of AI research and development processes that require regulatory or quasi regulatory instruments (in the absence of relevant legislation) and/or necessitate revisions in existing statutory frameworks.
The “how” question is related to the substance and form of AI regulations, in terms of their completeness, relevance, and accuracy. This argumentation is synonymous with the true and fair view concept applied in the accounting standards of financial statements.
The “when” question is concerned with the timeliness of the regulatory intervention. Policy makers ought to ensure that stringent rules do not hinder or delay the advancement of technological innovations.
The “where” question is meant to identify the context where mandatory regulations or the introduction of soft laws, including non-legally binding principles and guidelines are/are not required.
Future researchers are expected to investigate further these four questions in more depth and breadth. This research indicated that most contributions on AI governance were discursive in nature and/or involved literature reviews. Hence, there is scope for academic colleagues to conduct primary research activities and to utilize different research designs, methodologies and sampling frames to better understand the implications of planning, organizing, implementing and monitoring AI governance frameworks, in diverse contexts.
The full article is also available here: https://www.researchgate.net/publication/372412209_Artificial_intelligence_governance_Ethical_considerations_and_implications_for_social_responsibility
Substantia Mea 
You must be logged in to post a comment.