EH

Overview of AI Ethics - Detailed Study Notes

Introduction

  • AI ethics defined as the study of ethical issues in AI, focusing on how AI affects privacy, fairness, employment, safety, security, and social order.
  • AI technologies span autonomous driving, medical care, media, finance, industrial robots, and internet services; their widespread deployment improves efficiency but also poses risks.
  • Common ethical concerns: privacy leakage, discrimination, unemployment, security risks, and potential misuse by criminals.
  • Purpose of the article: provide a comprehensive overview of AI ethics through (i) ethical risks and issues, (ii) ethical guidelines and principles from various bodies, (iii) approaches to address ethical issues, (iv) methods to evaluate ethics in AI, and (v) challenges and future directions.
  • Impact statement underscores AI ethics as a crucial, cross-sector topic for researchers, practitioners, industry, government, and individuals.
  • Article timeline and attributions: manuscript received in 2021, revised 2022, accepted 2022, published 2022; current version 2023-07-21; funding sources and affiliations listed; supplementary material available online.
  • Index terms: Artificial intelligence (AI), AI ethics, ethical issue, ethical theory, ethical principle.

Scope and Methodology

  • Scope: four aspects are covered to form a systematic overview of AI ethics:
    • Ethical issues and risks of AI
    • Ethical guidelines and principles for AI
    • Approaches to addressing ethical issues in AI (ethical, technological, legal)
    • Methods for evaluating the ethics of AI
  • Methodology overview:
    • Literature sources include academic papers, organizational reports, government grey literature, and news articles.
    • Two-phase literature search strategy:
    • Phase 1: keyword-based search on multiple databases (Google Scholar, Web of Science, IEEE Xplore, ACM DL, Science Direct, Springer Link, arXiv, Google) using terms related to ethics, responsibility, trust, transparency, fairness, robustness, privacy, safety, sustainability, etc., coupled with AI terms (AI, ML, intelligent system).
    • Phase 2: examine related works and works by authors found in phase 1.
    • Guidelines: only English-language AI ethics guidelines; full list with URLs provided in Supplementary Materials.
  • The article aims to inform researchers, engineers, practitioners, and stakeholders with background, domain knowledge, and a bird’s-eye view to guide ethical AI practice.

Ethical Issues and Risks of AI

  • The section surveys existing categorizations of AI ethics issues and proposes an integrated, three-level framework.
  • Four existing categorizations are reviewed, then a new three-level categorization is presented.
  • A) Review of Categorizations of AI Ethical Issues (four approaches) 1) Based on Features of AI, Human Factors, and Social Impact (from governmental/academic sources)
    • AI features: Transparency/black-box problem in ML, data security and privacy risks, autonomy/intentionality/responsibility of autonomous systems.
    • Human factors: Accountability (the problem of many hands), Ethical Standards, Human Rights Laws (variation across countries/regions).
    • Social impact: Automation and job replacement; Accessibility of AI benefits; Democracy and civil rights concerns due to biased or opaque AI.
    • Example issues cited: lack of transparency in ML; data leakage; autonomous systems with high autonomy; civil rights concerns; public trust.
      2) Based on Vulnerabilities of AI and Human
    • AI vulnerabilities: data-hungry ML; garbage in, garbage out; faulty algorithms (underfitting/overfitting); deep learning as a black box affecting explainability and trust.
    • Human vulnerabilities: abuse of AI (facial recognition misuse, deepfake risks); job replacement concerns; ethical issues with robotic companions.
      3) Based on Algorithm, Data, Application, and Long-Term/Indirect Risks
    • Algorithm-related issues: algorithm security (model leakage, parameter tampering), explainability, algorithmic decision dilemmas (e.g., autonomous vehicle trade-offs).
    • Data-related issues: privacy protection, processing of personal/sensitive information, deidentification and data synthesis challenges.
    • Application-related issues: algorithmic discrimination and misuse/abuse of algorithms.
    • Long-term/indirect risks: employment impact, ownership of AI outputs, competition/monopoly concerns, responsibility for harms.
      4) Based on Deployment of AI
    • Impacts mapped to society, human psychology, financial systems, legal systems, environment, and trust.
    • Society: labor market disruptions, inequality, privacy and dignity, democracy risks, bias, polarization.
    • Human psychology: shifts in relationships and personhood concerns for AI agents.
    • Financial system: automation in trading; risk of destabilization or unpredictable AI-driven decisions.
    • Legal system: implications for criminal law, tort law, and liability when AI systems are involved.
    • Environment: resource use, pollution, energy consumption, sustainability concerns.
    • Trust: fairness, transparency, accountability, regulation; public trust as a core requirement.
  • B) Our Proposed Categorization: AI ethical issues at individual, societal, and environmental levels
    • Rationale: prior schemes miss environment or other aspects; need a tri-level framework to cover human-level, societal, and ecological impacts.
    • Categorizations:
      1) Individual level: safety, privacy, autonomy, human dignity. Includes risks from autonomous systems and data privacy concerns; autonomy erosion; risks to dignity (e.g., lethal autonomous weapons).
      2) Societal level: fairness/justice, responsibility/accountability, transparency, surveillance/datafication, controllability, democracy and civil rights, job disruption, human relationships.
      3) Environmental level: resource consumption (rare earths, metals), pollution and waste from hardware and devices, energy consumption, and sustainability of AI systems.
  • C) Key Ethical Issues Associated with Each Stage of the AI System Lifecycle
    • Lifecycle stages: business analysis, data engineering, ML modeling, deployment, operation and monitoring.
    • Business analysis: define problem and business metrics (including model performance and business KPIs).
    • Data engineering: data collection, labeling, cleaning, structuring, feature engineering.
    • ML modeling: algorithm design/selection, training, evaluation.
    • Deployment: making the model available to systems/users.
    • Operation/monitoring: ongoing performance evaluation and impact assessment.
    • A map links lifecycle stages to likely ethical issues (Table II in the paper).

Ethical Guidelines and Principles for AI

  • The field has produced a global landscape of guidelines and principles since 2015; the article aggregates 146 guidelines from governments, industry, academia, and international bodies.
  • The rest of the article (and supplementary materials) details the guidelines and the principles they promote.
  • A) Guidelines for AI Ethics
    • Earlier synthesis (Jobin et al., 2019) identified five core principles across 84 guidelines: transparency, justice/fairness, nonmaleficence, responsibility, and privacy.
    • UNESCO adopted the UNESCO Recommendation on the Ethics of Artificial Intelligence (2021) as a global agreement.
    • This article updates Jobin et al. with additional guidelines and presents 146 AI ethics guidelines in total.
    • A full list of collected guidelines is in the Supplementary Materials; the paper also reports a yearly distribution (2015–2021) with Table III showing the number of guidelines issued per year; 2018 had the largest count (53 guidelines, about 36.3% of total).
    • Table IV lists the number of guidelines issued by each country/region; Figure 3 shows the distribution by issuer type (government, industry, academia, other).
    • The core principles identified across guidelines (and elaborated in the paper) include 11 principles:
      1) Transparency
      2) Fairness & Justice
      3) Responsibility and Accountability
      4) Nonmaleficence
      5) Privacy
      6) Beneficence
      7) Freedom and Autonomy
      8) Solidarity
      9) Sustainability
      10) Trust
      11) Dignity
  • B) Principles for AI Ethics (detailed discussion of each principle)
    • 1) Transparency: involves transparency of the AI technology and the development/adoption process; interpretability, explainability, and justifiability of design and outcomes.
    • 2) Fairness & Justice: AI should avoid discrimination/bias; aim for just outcomes across individuals and groups.
    • 3) Responsibility and Accountability: AI must be auditable; designers, developers, owners, and operators should be responsible for AI behaviors and harms; mechanisms for accountability are needed.
    • 4) Nonmaleficence: do no harm; ensure safety and security to prevent malicious use and harm to humans.
    • 5) Privacy: protect privacy and data protection; govern data properly; protect data from theft; inform stakeholders about breaches.
    • 6) Beneficence: AI should do good and benefit humanity; align objectives with global challenges (e.g., food security, pollution, contagions).
    • 7) Freedom and Autonomy: preserve human decision-making authority; balance autonomy between humans and AI.
    • 8) Solidarity: AI should promote social security, cohesion, and maintain relational bonds across generations.
    • 9) Sustainability: AI development should be environmentally sustainable and avoid harm to ecosystems; balance AI progress with planetary health.
    • 10) Trust: trust in AI requires fairness, transparency, accountability, and regulation; trustworthy AI spans the entire socio-technical system.
    • 11) Dignity: AI should respect human dignity; avoid diminishing intrinsic human value; protect physical and mental integrity and identity.

Approaches to Address Ethical Issues in AI

  • The article emphasizes multidisciplinary approaches beyond pure technology, combining ethical, technological, and legal methods.
  • A) Ethical Approaches: Implementing Ethics in AI
    • Ethical theories provide a normative basis for AI ethics; three major normative branches are discussed: virtue ethics, deontological ethics, and consequentialism.
    • 1) Ethical Theories: metaethics, normative ethics, and applied ethics; normative branches include virtue, deontological, and consequentialist ethics; these theories guide what counts as morally right or wrong in AI behavior.
    • 2) Approaches for Implementing Ethics in AI (top-down, bottom-up, hybrid)
    • Top-down: encode explicit ethical theories/rules into AI systems; advantages include predictability and interpretability; disadvantages include inflexibility in dynamic environments.
    • Bottom-up: agents learn morality from experience and environment; advantages include adaptability; disadvantages include data requirements and potential inconsistency.
    • Hybrid: combine top-down rules with bottom-up learning to leverage strengths of both approaches; challenges include effective integration.
    • Remarks emphasize differences in interpretability, adaptability, and guidance, with a comparison summarized in Table VI.
  • B) Technological Approaches
    • The field is still in early stages for technological solutions addressing AI ethics.
    • Key topics include explainable AI (XAI) to address transparency; fair ML to reduce bias; safe/robust AI; privacy-preserving methods like differential privacy and federated learning; and multi-principle integration remains challenging.
    • The section notes that major venues (ACM FAccT, AAAI/AIES, IJCAI-ECAI) reflect growing interest, but practical, cross-principle integration is still limited and under development.
    • Table II in Supplementary Materials provides representative topics related to the five core principles.
  • C) Legal Approaches: Legislation and Regulation
    • The legal landscape has grown to include several landmark laws and regulations:
    • 2016: EU General Data Protection Regulation (GDPR)
    • 2017: US SELF DRIVE Act (Safety of automated vehicles)
    • 2018: Brazil LGPD (General Data Protection Law)
    • 2021: EU AI Act (regulatory framework for AI across the market)

Methods to Evaluate Ethical AI

  • The article emphasizes the importance of evaluating AI ethics before deployment, noting that this area has often been overlooked.
  • A) Testing
    • Moral Turing Test (MTT): extends the classic Turing Test to assess machine morality by restricting conversations to moral questions; a machine passes if humans cannot reliably distinguish it from a moral human counterpart.
    • Comparative MTT (cMTT): uses balanced pairs of actual moral actions by a human and a machine to assess which is more moral; acknowledges limitations such as consistency of machine behavior.
    • Expert vs. nonexpert tests: evaluators can be subject-matter experts in normative ethics or laypeople evaluating based on folk morality.
  • B) Verification
    • Formal verification treats AI behavior as a system S in an environment E with a property Φ to verify; outputs Yes/No, with counterexamples when No; emphasizes design verification to ensure alignment with ethical goals.
    • Arnold and Scheutz critique MTT and advocate for design verification as a core approach to ethical AI.
  • C) Standards
    • Various professional and organizational codes guide ethical conduct in AI practice:
    • 2014: ASC Professional Code of Conduct (Australian Computer Society)
    • 2018: ACM Code of Ethics and Professional Conduct (updated)
    • IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems (IEEE P7000 series)
    • ISO/IEC JTC 1/SC 42: AI standardization efforts (foundational standards, trust, governance, etc.)
    • These standards aim to complement guidelines with concrete, auditable expectations; however, there is a gap between high-level principles and practical implementation, especially for smaller organizations.

Challenges and Future Perspectives

  • The field is evolving, with several challenges identified across guidelines, implementation, technology, evaluation, and cross-cultural coordination.
  • A) Challenges in AI Ethical Guidelines and Principles
    • No universally adopted guideline; diverse principles across organizations, sectors, and regions; consensus on fundamental principles is lacking.
    • Application-specific principles may vary; need a basic set of common principles that can be adapted to domain contexts.
  • B) Challenges in Implementing Ethics in AI
    • Virtue ethics in practice: motives behind AI actions are often opaque; selecting which virtues to align with and how to measure them is difficult.
    • Deontological ethics in practice: choosing which duties to implement; conflicts among rules; prioritization of rules is challenging.
    • Consequentialist ethics in practice: uncertainty about consequences and quantifying utilities; lack of transparency complicates outcome prediction.
    • Coordination among different ethical standards across cultures and organizations; the goal is not necessarily full uniformity but effective coordination.
  • C) Challenges in Developing Technological Approaches
    • Integration of multiple ethical principles is hard; current research tends to focus on single principles (e.g., XAI for transparency, Fair ML for fairness).
    • Balancing multiple ethical requirements within a single system is complex and may involve trade-offs.
  • D) Challenges in Evaluating Ethics in AI
    • Ethics are qualitative and context-dependent; substantial subjectivity in evaluation; requires domain-specific benchmarks and multi-stakeholder input.
  • E) Future Perspectives
    • Multitheory approach: AI systems should be capable of switching between normative ethical theories depending on context, or combining domain-specific ethics with normative theories.
    • Domain-specific alignment: combining normative ethics with expert domain ethics to gain user acceptance and practical relevance.
    • Technological development guided by ethical guidelines and principles; need for integrated evaluation methods spanning multiple principles.
    • Development of domain-specific, comprehensive benchmarks for moral testing, especially for critical domains like autonomous vehicles and health care.
    • A combined approach that pairs normative ethics with evolutionary ethics to capture innate moral tendencies and adaptive, learning-based improvements over time.

Conclusion

  • AI ethics is a broad and rapidly evolving field with significant challenges in guidelines, implementation, evaluation, and cross-domain coordination.
  • A joint effort across AI scientists, philosophers, users, and policymakers is required to design ethical AI that can responsibly shape the future.
  • The article provides a comprehensive foundation, summarizing ethical risks, global guidelines, ethical approaches, evaluation methods, and future research questions to guide researchers and practitioners.

Supplementary Considerations and Notable References

  • Notable cases illustrating ethical risks include: autonomous vehicle accidents, Tay.ai incidents, CEO impersonation fraud, bias and discrimination in hiring or criminal justice, privacy breaches, and data misuse in surveillance contexts.
  • The article references important reviews and surveys on bias in ML, safety in RL, privacy in deep learning, explainable AI, and bibliometric studies of AI ethics trends.
  • Key equations, data points, and numbers mentioned in the article (presented with LaTeX formatting for emphasis):
    • Guidelines issued over time: ext{total guidelines} = 146
    • Largest annual guideline release: 53 in 2018 (36.3% of total)
    • Notable carbon footprint reference: training a Transformer model can have a carbon footprint that is approximately 5 times that of an average car over its lifetime (quoted in the paper as a comparison point)
    • Notable monetary example: criminals used AI to impersonate a CEO’s voice to request a fraudulent transfer of 243{,}000
  • These figures illustrate the scale of global guideline development, environmental considerations of AI, and real-world security risks discussed in the article.

Appendix: Key Tables and Figures Mentioned

  • Fig. 1: Topics covered in the article and the links between them.
  • Fig. 2: The proposed three-level categorization of AI ethics issues (individual, societal, environmental).
  • Fig. 3: Percentage of guidelines released by different types of issuers (government, industry, academia, other).
  • Table I: Review of categorizations of AI ethical issues and the authors’ proposed categorization.
  • Table II: Ethical considerations mapped along the AI lifecycle stages (business analysis, data engineering, ML modeling, deployment, operation/monitoring).
  • Table III: Number of documents issued each year from 2015 to 2021.
  • Table IV: Number of guidelines issued by country/region.
  • Table V: Comparison of the three normative ethical theories (virtue, deontological, consequentialist).
  • Table VI: Features of the three approaches for implementing ethics in AI (top-down, bottom-up, hybrid).
  • Supplementary Materials: Extended tables with list of guidelines, full principle taxonomy, and detailed references.