Source-linked AI summary
The Ethics of AI Ethics -- An Evaluation of Guidelines
Thilo Hagendorff
TL;DR
AI ethics guidelines have proliferated, but the paper asks whether their principles affect AI decision-making and practice. It analyzes and compares 21 major guidelines with concrete AI research and development practices. The paper finds that guidelines often have little practical effect, omit important ethical concerns, and require both more technical specificity and greater attention to social and personal responsibility.
Problem
AI ethics has many normative guidelines, but their tangible implementation and effect on AI research, development, application, and decision-making remain limited.
Method
The paper analyzes and compares 21 major AI ethics guidelines, examines their omissions, and compares their principles with concrete AI research and development practices.
Results
AI ethics guidelines often have little practical effect, while ethically motivated technical efforts address some issues and many other concerns remain omitted.
Takeaways & Limitations
Effective AI ethics requires both tangible bridges between abstract values and technical implementations and a situation-sensitive focus on knowledge, autonomy, and self-responsibility.
Takeaways & Limitations
The paper cautions that system-theoretical interpretations of economic behavior in the technology industry apply only at a macro level and must not be generalized.
Abstract
from arXiv · showhide
Current advances in research, development and application of artificial intelligence (AI) systems have yielded a far-reaching discourse on AI ethics. In consequence, a number of ethics guidelines have been released in recent years. These guidelines comprise normative principles and recommendations aimed to harness the "disruptive" potentials of new AI technologies. Designed as a comprehensive evaluation, this paper analyzes and compares these guidelines highlighting overlaps but also omissions. As a result, I give a detailed overview of the field of AI ethics. Finally, I also examine to what extent the respective ethical principles and values are implemented in the practice of research, development and application of AI systems - and how the effectiveness in the demands of AI ethics can be improved.
1 Introduction
The paper examines whether AI ethics guidelines influence decision-making and practice, finding that they generally do not. It analyzes major guidelines, compares their omissions, assesses implementation, and proposes ways to make AI ethics actionable.
- 1 Introduction: Most AI ethics guidelines have little or no actual impact on human decision-making in AI and machine learning.The paper attributes this weakness partly to ethics’ lack of effective reinforcement mechanisms.
- 1 Introduction: The paper analyzes 21 major AI ethics guidelines and identifies issues they omit.The guidelines are compared to determine their coverage and omissions.
- 1 Introduction: It compares guideline principles with concrete research and development practices to examine whether ethical objectives are implemented.The analysis focuses on whether principles have an effect in practice, rather than remaining good intentions.
- 1 Introduction: The paper proposes transforming AI ethics from a discursive phenomenon into concrete directions for action.This includes addressing the difficulty of deriving technological implementations from abstract values such as justice, transparency, and human-centeredness.
2 Guidelines in AI ethics
The paper situates its analysis within comprehensive AI ethics guidelines, systematically compiles 21 major documents, and compares their ethical coverage. The guidelines repeatedly emphasize values that are technically operationalizable, while omitting several broader and speculative concerns.
- 2 Guidelines in AI ethics: The study focuses on comprehensive AI ethics guidelines rather than autonomous-machine implementation, meta-studies, trolley-problem analyses, or specific-problem reflections.It examines compilations intended to map and categorize normative claims across AI ethics.
- 2.1 Method: The selection of 21 major guidelines used a two-phase literature analysis covering recent documents and excluding purely national-context materials.Searches covered Google, Google Scholar, Web of Science, ACM Digital Library, arXiv, and SSRN; selection prioritized comprehensive normative coverage over document detail.
- 2.2 Multiple entries: Accountability, privacy, and fairness appear in nearly 80% of guidelines and form recurring minimum requirements for ethically sound AI.These topics are also among those for which technical solutions or tools have been developed.
- 2.2 Multiple entries: Accountability, interpretability, privacy, justice, transparency, robustness, and safety tend to be framed as technical solutions because they are mathematically operationalizable.This pattern is associated with technical communities such as FAT ML and XAI.
- 2.2 Multiple entries: Women constituted 37.1% of indicated guideline authors, falling to 31.3% when the three AI Now Reports were excluded and to 7.7% in FAT ML guidelines.The analysis links this distribution to the limited presence of care-oriented, welfare, social-responsibility, and ecological perspectives.
- 2.3 Omissions: Guidelines omit or rarely discuss malevolent AGI, existential threats, machine consciousness, robot ethics, and trolley problems.Robot-ethics guidelines were intentionally excluded, while the examined guidelines gave little attention to trolley problems and related autonomous-vehicle questions.
3 AI in practice
AI’s practical use is shaped by harmful applications, weak guideline enforcement, and organizational incentives that often diverge from ethical principles. Although privacy, fairness, and explainability have seen technical progress, many ethical objectives remain difficult to assess or substantially underachieved.
- AI is used in military systems, propaganda, surveillance, social control, and other contexts involving unwanted or directly harmful effects.
- AI competition intensifies in-group and out-group thinking, making ethical concerns about social division especially relevant to the field’s broader context.
- Ethical guidelines have little influence on professionals’ behavior because compliance is voluntary, enforcement mechanisms are weak, and ethical concerns are rarely empowered organizationally.
- AI research, business, and industry operate through distinct institutional codes and values, limiting ethical intervention, although employee protests show that deviations from economic logic occur.
- Technical progress is most visible in privacy, fairness, and explainability, including cryptographic and differential-privacy methods for limiting AI systems’ access to data.
- Guideline objectives are difficult to evaluate for safety, cybersecurity, employment, human oversight, and related issues, while sustainability, diversity, autonomy, and social cohesion are markedly underachieved.
4 Advances in AI ethics
The paper argues that broad AI ethics principles need translation into concrete technical and organizational practices. It proposes more context-sensitive guidance, virtue ethics, stronger collaboration, distributed responsibility, and institutional support.
- 4.1 Technical instructions: Because “AI” covers heterogeneous technologies, ethics should examine data practices, algorithm design, code, and training-set selection at a more concrete microethical level.
- 4.1 Technical instructions: Standardized datasheets for training datasets can document their properties, origins, composition, collection, and preprocessing for practitioners.
- Broad AI guidelines remain distant from the diverse scientific, technical, and economic practices they aim to govern.
- 4.2 Virtue ethics: The paper recommends augmenting deontological ethics with virtue ethics focused on values, character, responsibility, courage, autonomy, and self-responsibility.
- 4.2 Virtue ethics: AI ethics requires closer collaboration between ethicists and technical researchers, including mutual openness to ethical and technical knowledge.
- 4.2 Virtue ethics: Responsibility should extend across everyone involved in AI data, engineering, and applications, including consideration of alternative development paths or refraining from unethical tasks.
- Institutional support should include legal frameworks, independent audits, complaint and compensation mechanisms, and expanded university curricula in technology ethics.
5 Conclusion
The paper concludes that AI ethics is often ineffective because ethical commitments lack reinforcement and economic incentives can override them. It therefore calls for moving beyond checklist guidelines toward context-sensitive responsibility that connects technical implementation with social and personality-related concerns.
- AI ethics often fails because deviations from ethical codes have few consequences, while institutionalized ethics may function mainly as marketing.
- Economic incentives can override commitment to ethical principles, leaving AI purposes misaligned with societal values and fundamental rights.
- Ethically motivated technical fixes address some problems, but guidelines omit wider concerns including political abuse, social cohesion, diversity, and hidden social and ecological costs.
- AI ethics should move beyond checkbox guidelines toward situation-sensitive practice grounded in virtues, knowledge, responsible autonomy, and empathy.
- AI ethics must build tangible bridges between abstract values and technical implementations while also addressing social and personality-related aspects.