A preliminary literature review finds a growing number of ethical AI frameworks for the design and governance of AI systems being developed and a recent focus on the operationalization of AI ethical principles (guidelines) across societal sectors and industries.
- Introduction
- Salient AI principles
- AI guidelines
- Criticisms of guidelines
- Application frameworks
- Issues/challenges within ethical AI
- Ethical AI paradigms
- Towards a pragmatic design approach: AI-STEI-KW
You may also be interested in Ethical assessment of teaching ethical hacking.
Introduction
AI ethical principles can be discussed under the broad umbrella term of AI ethics. In comparison, the domain of ethical AI is concerned both with an AI code of ethics as well as the operationalization of ethical principles within AI systems.
AI ethics refers to moral principles and social values intended to guide the development and use of AI systems. It is a normative approach to technology regulation that draws on salient philosophical perspectives, mainly consequentialist (e.g., utilitarianism) and deontological (e.g., duty), to determine what constitutes fair and responsible technology use.
“AI ethics involves the ethics of AI, which studies the ethical theories, guidelines, policies, principles, rules, and regulations related to AI” (Siau & Wang, 2020). AI ethics is concerned with the ethical and social impacts of technology on society and individuals within sociotechnical systems. For Christoforaki and Beyan (2022), AI ethics is “a multidisciplinary domain that has philosophical, legal, societal, and technological aspects.”
Salient AI principles
*Recent surveys ([57,58,59,60,61], cited in Christoforaki & Beyan, 2022) have identified common principles or themes that run through the AI guidelines landscape as: transparency, justice, fairness, non-maleficence, responsibility, privacy, beneficence (in the terms of sustainability, well-being and common good), freedom, autonomy, trust, dignity, solidarity, accountability, auditability, safety and security, explainability, human control of technology, promotion of human values, humanity, collaboration, share, and AGI/ASI.
*A systematic review by Jobin et al. (2019) of 84 ethical guidelines identified 11 principles: 1) transparency; 2) justice and fairness; 3) nonmaleficence; 4) responsibility; 5) privacy; 6) beneficence; 7) freedom and autonomy; 8) trust; 9) dignity; 10) sustainability; and 11) solidarity. Jobin et al. (2019) found,
a wide divergence in how these principles were interpreted and in how the recommendations suggested (can) be applied. They do, however, note some evidence of convergence (by number of mentions) around five principles: 1) transparency; 2) justice and fairness; 3) nonmaleficence; 4) responsibility; and 5) privacy. (Peters, Vold, Robinson, & Calvo, 2020)
*Floridi et al.’s (2018) Ai4People framework incorporates the classical four principles of biomedical ethics for “safeguarding of individuals in instances where technological practices affect their interests and wellbeing” (Beauchamp & Childress, 2019, p. 9): Respect for Autonomy, Beneficence, Nonmaleficence, and Justice. Floridi et al. (2018) add explicability.
*The main criticism levelled at the basic principles is that they are too high level to be immediately implementable within industries and organizations. Broad principles “do not offer specific practices to apply ethics at each stage of the AI/ML pipeline, and often fail to be actioned in governmental policy [64,65,66]” (Christoforaki & Beyan, 2022).
AI guidelines
Spearheaded by government and civil society organizations, mainly from the west, AI regulation efforts began around 2010 with an emphasis on establishing ethical principles for the design, development, and use of AI. According to the NGO AlgorithmWatch (2022), which maintains a global inventory of AI ethical guidelines, there are currently 173 frameworks (ethical guidelines), many of which were released within the last five years (Christoforaki & Beyan, 2022).
Of the 173 guidelines in the AlgorithmsWatch inventory, “115 are non-enforceable recommendations, while regarding the 41 which come from the private sector, 15 are recommendations and 22 are voluntary commitments leaving only 4 characterized as binding agreements, i.e., having the provision for means to sanction non-compliance” (Christoforaki & Beyan, 2022).
Recent AI ethical guidelines include “the OECD Principles in 2019, the Montreal Declaration in 2017, the U.K. House of Lords report ‘AI in the U.K.: ready willing and able?’ in 2018, the European Commission High-Level Expert Group (HLEG) on AI in 2018, and the Beijing AI Principles in 2019” (Peters, Vold, Robinson, & Calvo, 2020).
Criticisms of guidelines
*Agreeing on a set of general principles is difficult because practitioners and other stakeholders might interpret the principles differently, depending on their geographical and cultural backgrounds and ways of working. Practitioners belong to diverse scientific domains, and have access to different resources, infrastructures, and funding (Christoforaki & Beyan, 2022).
For example, Jobin et al. [57] and Zeng et al. [61] include transparency and explainability in the general notion of transparency, while Hagendorff [58] examined them as different issues; Floridi et al. (2018) included accountability in explicability [60], while for Hagendorff it was a separate issue [58], and Jobin et al. [57] and Zeng et al. [61] included it in responsibility. (Christoforaki & Beyan, 2022)
*The mere existence of ethical guidelines does not ensure or necessarily directly correlates with the implementation of the principles.
A behavioral ethics study involving 63 software engineering students and 105 professional software developers revealed that the decision-making by the participants that were explicitly instructed to consider the Association for Computing Machinery (ACM) code of ethics, was not statistically significantly different when compared with a control group. (Christoforaki & Beyan, 2022)
Application frameworks
A growing number of approaches are being developed to translate abstract principles into actionable practice. Several of these application frameworks integrate various technical solutions to operationalize the ethical principles. But few are considered comprehensive to the point of elaborating concrete procedures for technical implementation.
Ethical AI application frameworks “usually consist of defining a framework within which the ethical principles or values are operationalized with regard to the various stages of an AI project process model—usually CRISP-DM (cross industry standard process for Data Mining) or similar” (Christoforaki & Beyan, 2022).
*One example of a suggestion that would operationalize abstract principles such as transparency and accountability is “work by Gebru et al. [13] on ‘datasheets for datasets’ in which they advocate for clear documentation for datasets that records ‘motivation, composition, collection process, recommended uses, and so on'” (Peters, Vold, Robinson, & Calvo, 2020).
*The Alan Turing Institute’s guide for the responsible design and implementation of AI systems in the public sector proposes “an ethical platform for the responsible delivery of AI projects, the purpose of which is to safeguard and enable the justifiability of both the AI project and its product” (Christoforaki & Beyan, 2022). The platform consists of three building blocks: 1) The value: respect, connect, care, and protect; 2) the principles: fairness, accountability, sustainability, and transparency; and 3) a process-based governance framework across the AI system design and implementation workflow processes ” – CRISP-DM and other related workflow models (Christoforaki & Beyan, 2022).
*The High-Level Expert Group on AI (AI HLEG) set up by the European Commission, which Statistics Canada is basing its Framework for Responsible Machine Learning Processes on, proposes a framework for Trustworthy AI, “a concept that comprises lawful, ethical, and robust (i.e., secure and reliable) AI” (Christoforaki & Beyan, 2022).
Trustworthy AI is based on a set of initial principles, namely respect for human autonomy, prevention of harm, fairness, and explicability. These principles, in turn, define seven requirements: (1) human agency and oversight; (2) technical robustness and safety; (3) privacy and data governance; (4) transparency; (5) diversity, non-discrimination, and fairness; (6) environmental and societal well-being; and (7) accountability, which are subsequently implemented by technical and non-technical methods. The technical methods comprise of: system architectures which encompass procedures and/or constraints on procedures which implement trustworthy AI; ethics and rule of law by design, i.e., methods to ensure that abstract principles are translated into specific implementation decisions from the very start; explanation methods; testing and validating; and finally, quality of service indicators, such as measures to evaluate the testing and training of algorithms, functionality, performance, usability, reliability, security, and maintainability. The outcome for each one of the requirements should be continuously assessed during a system’s life cycle [77]. (Christoforaki & Beyan, 2022)
*While most of the guidelines refer to abstract principles (Hagendorff, 2020) “they provide no, or very few notes on technical implementation” (Christoforaki & Beyan, 2022). Only two of the 22 guidelines documents that Hagendorff (2020) surveyed contain notes on technical implementation, namely, 1) the AI principles of the European Commission’s High-Level Expert Group on AI, and 2) Floridi et al.’s (2018) AI4People—An Ethical Framework for a Good AI Society (Christoforaki & Beyan, 2022).
*Steven Umbrello and Ibo van de Poel (2021) propose a modified VSD approach that incorporates the AI for Social Good (AI4SG) principles by Floridi, Cowls, King, and Taddeo (2020) to translate abstract philosophical values into tangible design requirements.
Issues/challenges within ethical AI
*Big Data bias/fairness: detecting and mitigating bias throughout the AI data life cycle.
Bias can enter in each step of an AI/ML pipeline. Suresh and Guttag [42] offer a comprehensive description framework that identifies potential sources of harm. It spans from the creation of the datasets (data collection and preparation), to model building and implementation (model development, evaluation, post-processing, and deployment). The creation of the datasets can be adversely impacted by (a) historical bias, which incorporates already existing bias in the world; (b) representational bias, i.e., underrepresentation of specific populations during the sampling process; and (c) measurement bias, where the choice of features and labels that are assigned to data that work as proxies for the desired quantities is poor. These probably already problematic datasets are fed to a model which may be impacted by (a) aggregation bias that is created by combining heterogeneous datasets during model construction; (b) evaluation bias, i.e., when the benchmark population does not correspond to real-world cases; and (c) deployment bias, where the outcomes of the system are interpreted and used in an inappropriate manner. (Christoforaki & Beyan, 2022)
*Privacy issues and solutions for machine learning.
Liu et al. (2021) offered a comprehensive survey on AI/ML and privacy, addressing three specific cases and identifying the key challenges. Liu et al. examined AI as either “the object of privacy concerns or the tool to avert or instigate privacy attacks.” “Where the objective is to ensure the privacy of either the training data or the algorithm, they classify the attacks and protection schemes and assess the existing solutions. Regarding the other two cases, namely AI aided privacy protection and AI-based privacy attack, they conclude that while the first one is gaining momentum, the research about the latter is in infancy but it is expected to have great future development. (Christoforaki & Beyan, 2022)
Ethical AI paradigms
Responsible AI, Trustworthy AI, Reliable AI, Explainable AI, Accountable AI, Transparent AI, Fair AI, Secure AI, etc.
*eXplainable AI (XAI). XAI aims to produce explainable models or effective individual explanations, thus enabling users to understand and trust the system (for a comprehensive overview, see [85]). While XAI is often connected to EU legislation -GDPR and “the right to explanation” [86], whether this includes a right to explanation from automated procedures is a debated issue [87]. (Christoforaki & Beyan, 2022)
Towards a pragmatic design approach: AI-STEI-KW
“Create a data and AI ethical risk framework that is tailored to your industry,” notes Blackman (2020).
In finance, it is important to think about how digital identities are determined and how international transactions can be ethically safe. In health care there will need to be extra protections built around privacy, particularly as AI enables the development of precision medicine. In the retail space, where recommendation engines loom large, it is important to develop methods to detect and mitigate associative bias, where recommendations flow from stereotypical and sometimes offensive associations with various populations. (Blackman, 2020)
The first step in the deployment of ethical AI within an organizations is establishing a committee or taskforce in charge of designing and implementing ethical AI. Within an organization’s industry and within the specific intended application of AI, risk scenarios (including liabilities) and improvement/automation goals should be articulated. Next, a suitable ethical AI framework (as a risk management framework) should be selected as a broad framework to guide the design and implementation process. STEI-KW is especially suitable for information security governance and data analysis tasks, including automated data collection and analysis. STEI-KW puts emphasis on the basic principles of efficiency, fairness, privacy, security, safety, and autonomy. STEI-KW takes a constructivist approach to knowledge management and collaborative design, directly engaging key stakeholder groups in a project’s lifecycle. One of the early objectives of the ethical AI taskforce will be the identification of key stakeholders who will consequently be involved in the design, execution, and assessment/oversight of the ethical AI project. The operationalization of basic principles at each stage of the data lifecycle will be done through close collaboration between data scientists/data engineers/data analysts and AI ethicists.
Components of STEI-KW as an AI risk management (ethical AI) framework
1. Ethical assessment of teaching ethical hacking
Table: The ethical case for teaching higher education students hacking skills
Table: Opportunities and risks of teaching students hacking skills
2. Canadian identity as an academic idea
Key references
AlgorithmWatch
“A non-profit research and advocacy organization that is committed to watch, unpack and analyze automated decision-making (ADM) systems and their impact on society.”
Blackman, R. (2020). A practical guide to building ethical AI. Harvard Bus. Rev, 15.
Christoforaki, M., & Beyan, O. (2022). AI Ethics—A Bird’s Eye View. Applied Sciences, 12(9), 4130.
Floridi, L. (Ed.). (2021). Ethics, governance, and policies in artificial intelligence. Springer.
Related content
A framework for understanding NLP
Canadian identity as an academic idea
Compliance frameworks and industry standards
Ethical assessment of teaching ethical hacking
Ethical decision-making theories: Introduction to normative ethics
The GRC approach to managing cybersecurity
The technoethics of Mario Bunge
Back to DTI Courses
