Source-linked AI summary
Artificial Intelligence: the global landscape of ethics guidelines
Anna Jobin, Marcello Ienca, Effy Vayena
TL;DR
Ethical AI lacks agreement on the requirements, standards, and practices needed for its realization. This paper maps and analyzes 84 guidelines, finding convergence around five principles alongside substantive divergence in their interpretation and implementation.
Problem
Developing a global AI agenda requires balancing cross-national harmonization with cultural diversity and moral pluralism.
Method
The study maps 84 eligible ethical-AI documents and analyzes them using reflective equilibrium.
Results
The analysis finds cross-stakeholder convergence on transparency, justice, non-maleficence, responsibility, and privacy, alongside substantive divergence.
Takeaways & Limitations
Developing ethical AI guidelines requires moving beyond principle formulation toward translation into practice.
Takeaways & Limitations
The retrieval process is inevitably less replicable and unbiased than a systematic database search.
Abstract
from arXiv · showhide
In the last five years, private companies, research institutions as well as public sector organisations have issued principles and guidelines for ethical AI, yet there is debate about both what constitutes "ethical AI" and which ethical requirements, technical standards and best practices are needed for its realization. To investigate whether a global agreement on these questions is emerging, we mapped and analyzed the current corpus of principles and guidelines on ethical AI. Our results reveal a global convergence emerging around five ethical principles (transparency, justice and fairness, non-maleficence, responsibility and privacy), with substantive divergence in relation to how these principles are interpreted; why they are deemed important; what issue, domain or actors they pertain to; and how they should be implemented. Our findings highlight the importance of integrating guideline-development efforts with substantive ethical analysis and adequate implementation strategies.
Introduction
AI’s transformative societal impact has intensified debate over the principles and values guiding its development and use. This study maps ethical-AI guidelines worldwide to assess convergence and divergence in principles and implementation requirements.
- Motivation: AI’s expanding impact across societal domains has prompted debate over the principles and values that should guide its development and use.Concerns include job displacement, misuse, weak accountability, bias, and threats to fairness.
- Motivation: Governments, companies, professional associations, and nonprofits have issued AI principles and guidance, reflecting both the need for ethical guidance and competing stakeholder priorities.Private-sector involvement has also been criticized as potentially framing social problems as technical or avoiding regulation.
- Research problem: The study asks whether diverse groups converge on what ethical AI should be and which principles determine AI development, or whether their differences can be reconciled.The paper treats both the composition of guidance-producing groups and the content of their guidance as important questions.
- Method and contribution: The authors conduct a scoping review to map the global landscape of ethical-AI principles and guidelines and assess convergence in principles and requirements for their realization.The analysis is intended to inform stakeholders advancing ethically responsible AI innovation.
Results
The corpus comprised 84 ethical AI documents, concentrated in recent publications and economically developed countries, with varied issuers and audiences. Eleven principles emerged, with convergence around five widely referenced principles but substantial divergence in their interpretation and implementation.
- Ethical principles: Eleven overarching ethical values and principles emerged, led by transparency, justice and fairness, non-maleficence, responsibility, and privacy.Other identified principles included beneficence, freedom and autonomy, trust, dignity, sustainability, and solidarity.
- Ethical principles: No single principle appeared across the entire corpus, but five principles were referenced in more than half of all sources.The five were transparency, justice and fairness, non-maleficence, responsibility, and privacy.
Justice, fairness, and equity
Ethical AI guidelines frame justice primarily through fairness, bias prevention, and nondiscrimination, while also addressing diversity, inclusion, equality, access, and redress. They pursue these aims through technical measures, transparency, testing and auditing, legal safeguards, oversight, stakeholder participation, and clear liability, alongside harm-prevention strategies centered on safety, security, privacy, and risk mitigation.
- Justice, fairness, and equity: Justice is primarily defined through fairness, prevention or mitigation of bias, and discrimination, with discrimination referenced significantly less by private-sector sources.
- Justice, fairness, and equity: Guidelines also emphasize diversity, inclusion, equality, appeal and redress rights, and fair access to AI, data, and its benefits.
- Justice, fairness, and equity: Justice is supported by accurate, complete, and diverse data, especially training data, alongside standards, normative encoding, transparency, testing, monitoring, and auditing.
- Justice, fairness, and equity: Broader implementation strategies include rule-of-law safeguards, appeal and remedy mechanisms, governmental oversight, interdisciplinary workforces, stakeholder inclusion, and attention to benefit distribution.
- Justice, fairness, and equity: Non-maleficence references outweigh beneficence by a factor of 1.5 and focus on safety, security, avoidance of foreseeable harm, and risks including discrimination, privacy violations, and bodily harm.Guidelines propose technical and governance measures across research, design, deployment, legislation, oversight, testing, monitoring, auditing, and liability attribution.
Responsibility and accountability
Responsibility and accountability are widely invoked but rarely defined, with disagreement over attribution, responsible actors, and whether AI itself can be accountable. Privacy is framed as both a value and a right, pursued through technical, research, awareness, and regulatory measures.
- Responsibility and accountability: Responsibility and accountability are widely referenced but rarely defined, with recommendations emphasizing integrity and clarifying responsibility and legal liability.Attribution may be established upfront, in contracts, or through remedies.
- Responsibility and accountability: Guidelines variously focus on the reasons and processes leading to harm, whistleblowing, diversity, and ethics education in STEM.These recommendations address both prevention and institutional culture.
- Responsibility and accountability: Accountability is assigned variously to developers, designers, institutions, or industry, while sources disagree over AI-like accountability versus exclusive human responsibility.Some sources maintain that humans should remain ultimately responsible for technological artifacts.
- Responsibility and accountability: Privacy is presented both as a value and a right, often linked to data protection and data security, and sometimes to freedom or trust.The concept is often left undefined.
- Responsibility and accountability: Privacy implementation spans technical solutions, further research, awareness, and regulatory approaches including legal compliance, certificates, and AI-specific laws.Examples include differential privacy, privacy by design, data minimization, and access control.
Beneficence
Beneficence is frequently mentioned but rarely defined, with guidelines differing over what promoting good means, who should benefit, and how AI should advance it.
- Beneficence: Beneficence is rarely defined, although exceptions invoke augmenting human senses, human well-being and flourishing, peace and happiness, socioeconomic opportunities, and economic prosperity.These interpretations frame benefit in terms of human capabilities, welfare, social conditions, and prosperity.
- Beneficence: Guidelines differ over beneficiaries, variously emphasizing customers, everyone, humanity, society, as many people as possible, sentient creatures, the planet, or the environment.Private-sector issuers tend to highlight customers, while other sources extend benefits across humans, nonhumans, and the environment.
- Beneficence: Proposed strategies include aligning AI with human values, advancing scientific understanding, minimizing power concentration and conflicts of interest, and using power for human rights.The guidelines connect beneficence to governance, knowledge, rights, and limiting concentrated power.
- Beneficence: Other implementation approaches include demonstrating beneficence through customer demand and feedback, and developing new metrics and measurements.These approaches seek evidence of benefit through stakeholder response and explicit evaluation tools.
Freedom and autonomy · Trust · Sustainability
The guidelines interpret freedom and autonomy through positive and negative freedoms, while linking trust to reliable, accountable, and sometimes explainable AI. Sustainability encompasses environmental protection, social equality, peace, durable systems, ecological efficiency, and accountability for job losses.
- Freedom and autonomy: Freedom and autonomy encompass expression, self-determination, privacy-protecting controls, empowerment, and autonomy, including freedoms to flourish, form relationships, withdraw consent, and choose technologies.They also include freedom from technological experimentation, manipulation, and surveillance.
- Freedom and autonomy: Freedom and autonomy are promoted through transparent, predictable AI, expanded knowledge, notice and consent, and restraint in collecting or disseminating data.Guidelines also emphasize preserving citizens’ options and knowledge.
- Trust: Trust guidelines address trustworthy AI research, technology, developers, organisations, design principles, and customers’ trust.Trust is justified as supporting organisational goals and enabling AI to fulfill its world changing potential.
- Trust: The guidelines both promote trust as indispensable and warn against excessive trust in AI, revealing disagreement about its appropriate role.Other facilitators include fairness certification, multi-stakeholder dialogue, and awareness of personal-data value.
- Trust: Trust-building measures include education, reliability, accountability, ongoing integrity monitoring, compliance tools, transparency, understandability, and explainability.One guideline instead recommends fulfilling public expectations rather than demanding understandability.
- Sustainability: Sustainability calls for AI that protects the environment, improves ecosystems and biodiversity, contributes to fairer societies, and promotes peace.Guidelines also envision sustainable systems, sustainable data processing, and insights that remain valid over time.
- Sustainability: Sustainable AI should be designed, deployed, and managed carefully to improve energy efficiency and minimize ecological footprints.Corporations are also asked to ensure accountability regarding potential job losses and treat challenges as opportunities.
Dignity
Existing guidelines largely leave dignity undefined, but commonly associate it with human rights, harm avoidance, and protecting human dignity from harmful AI impacts. They propose preserving dignity through developer respect, legislation, governance, and government-issued technical and methodological guidance.
- Dignity remains undefined in existing guidelines, except for the specification that it is a prerogative of humans but not robots.
- Guidelines frequently link dignity to human rights and avoiding harm, including forced acceptance, automated classification, and unknown human-AI interaction.
- AI should not diminish or destroy human dignity, but should respect, preserve, or even increase it.
- Dignity is believed to be preserved when respected by AI developers and promoted through legislation, governance initiatives, or government-issued technical and methodological guidelines.
Solidarity
Solidarity is primarily discussed in terms of AI’s implications for the labor market. Guidelines emphasize social protection, equitable distribution of AI’s benefits, protection of vulnerable groups, and avoiding data practices that promote radical individualism.
- Solidarity: Solidarity is mostly referenced in relation to AI’s implications for the labor market.
- Solidarity: Sources call for a strong social safety net and redistribution of AI’s benefits to protect social cohesion.
- Solidarity: Guidelines stress respecting potentially vulnerable persons and groups while warning that individual-focused data practices may undermine solidarity in favour of ‘radical individualism’.
Discussion
The discussion finds growing international and cross-sector engagement with ethical AI, alongside convergence around several principles but substantial divergence over their interpretation and implementation. It emphasizes unresolved tensions involving geographic representation, cultural pluralism, oversight, and translating principles into practice.
- Global engagement: Guidance documents for ethical AI are increasing rapidly in number and variety, reflecting growing international community involvement across public and private sectors.The nearly equivalent representation of public and private issuers indicates that ethical AI challenges concern both public entities and private enterprises.
- Global fairness and pluralism: Underrepresentation of Africa, South and Central America, and Central Asia suggests that MEDC countries shape the debate more than others, raising concerns about local knowledge and moral pluralism.A global agenda must balance cross-national harmonization with cultural diversity and moral pluralism, potentially through deliberative mechanisms.
- Principled convergence: Guidelines converge around transparency, justice and fairness, non-maleficence, responsibility, and privacy, although no single principle is endorsed by every guideline.Each of these five principles appears in more than half of all guidelines.
- Interpretive and implementation divergence: Substantive divergence persists in how principles are interpreted, why they matter, which issues, domains, or actors they concern, and how they should be implemented.Guidelines also leave unresolved which principles to prioritize, how to resolve conflicts, who should enforce oversight, and how institutions should comply.
- Practical tensions: Proposed measures can conflict, including calls for larger and more diverse datasets to reduce bias versus increased individual control over data for privacy and autonomy.Further tension appears between avoiding harm at all costs and balancing risks and benefits, whose evaluation depends on whose well-being actors optimize.
- Governance mechanisms: Independent Review Boards may be increasingly needed to assess AI applications in scientific research, but government and corporate applications may remain outside their oversight without expanded authority.The discussion also calls for greater governmental cooperation to harmonize and prioritize AI agendas, mediated by inter-governmental organizations.
Limitations
The study’s corpus and analysis are subject to limitations associated with gray-literature retrieval, search personalization, language bias, qualitative methods, and the rapidly changing AI-guidance landscape. The researchers mitigated these risks through protocol development, broad searching, independent coding, and continuous monitoring through April 23, 2019.
- Corpus retrieval: Gray-literature guidelines are less replicable and potentially less unbiased to retrieve than peer-reviewed literature in systematic databases.A discovery and eligibility protocol was developed and pilot-tested before data collection.
- Search bias: Personalized search-engine results may influence discovery, while broad keywords and inclusion criteria were used to mitigate this risk.The corpus may also be skewed toward English-language results.
- Analytic methods: The content analysis carries typical limitations of qualitative methods, including potential subjective bias.An inductive coding strategy was conducted independently by two reviewers to minimize subjectivity.
- Corpus completeness: Because AI guidance documents are published rapidly, documents may have appeared after the search was completed.The literature was continuously monitored during data analysis through April 23, 2019.
Methods
The study used a PRISMA-adapted scoping review to identify and synthesize ethical AI principles from gray literature. Researchers analyzed 84 eligible documents through multi-stage retrieval, qualitative coding, and ethics-informed consistency checks.
- Retrieval: Three search strategies—linkhubs, web search, and citation chaining—plus manual monitoring identified 84 eligible, non-duplicate documents containing ethical AI principles.The literature was monitored through April 23, 2019, and citation chaining continued until no additional relevant documents were found.
- Eligibility: Eligibility required policy documents in five languages, issued by institutional entities, explicitly addressing AI-related notions, and expressing a normative ethical stance.Included materials comprised principles, guidelines, and institutional reports from public and private sectors.
- Analysis: Two researchers independently analyzed the 84 sources through manual coding and code mapping in Nvivo for Mac v.11.4.The first coding cycle produced 3457 codes, including 1180 pertaining to ethical principles.
- Analysis: Researchers organized codes into ethical categories using qualitative metasynthesis, deductive normative literature, and consistency assessment by two ethics specialists.Thirteen emerging ethical categories were identified, with two merged because of semantic and thematic proximity.
- Analysis: Focused coding extracted category significance and frequency, while reflective equilibrium supported consistency between ethical literature, general principles, and particular policy judgments.Consistency was also maintained through deliberative mutual adjustment among the researchers’ interpretations.