The ethics of artificial intelligence addresses normative questions about the design, development, deployment, and governance of systems within Artificial intelligence and Machine learning, with emphasis on impacts on individuals, institutions, and society. International instruments frame core values—human rights, fairness, transparency, accountability, privacy, safety, and human oversight—and increasingly translate them into operational guidance and law. The OECD’s intergovernmental AI Principles (adopted in 2019, updated in 2024) articulate human‑rights‑respecting, trustworthy AI and corresponding policy recommendations for governments, forming the basis of the G20 AI Principles, according to the OECD AI Policy Observatory and related papers. (OECD AI principles; OECD 2019 working paper; OECD implementation reports and 2021 insights.)

Origins and global frameworks

• –UNESCO’s first global standard on AI ethics, the Recommendation on the Ethics of Artificial Intelligence, was adopted on 23 November 2021 and centers human rights, human dignity, transparency, accountability, human oversight, and sustainability, with detailed policy action areas for implementation. (UNESCO Legal Affairs page; UNESCO overview and values/principles summary (UNESCO dataviz).)
• –The European Commission’s High‑Level Expert Group issued Ethics Guidelines for Trustworthy AI (2019), defining lawful, ethical, and robust AI and seven key requirements used widely as a baseline. (European Commission, Trustworthy AI Guidelines.)
• –In health, the World Health Organization issued Ethics and Governance of AI for Health (2021) and subsequent guidance on large multimodal models (2024), emphasizing patient safety, inclusiveness, transparency, accountability, and proportionate regulation. (WHO 2021 guidance; WHO LMM guidance 2024.)

Regulatory and policy developments

• –The European Union adopted the AI Act (Regulation (EU) 2024/1689), published in the Official Journal on 12 July 2024, entering into force on 1 August 2024; prohibitions apply from 2 February 2025, many GPAI obligations and governance provisions apply from 2 August 2025, and most other requirements apply from 2 August 2026. (EUR‑Lex AI Act; timeline summary (AI Act timeline).)
• –In the United States, the White House released the “Blueprint for an AI Bill of Rights” (October 2022) identifying five protections—safe and effective systems, algorithmic discrimination protections, data privacy, notice and explanation, and human alternatives—and an Executive Order (October 30, 2023) directing agencies to advance AI safety, civil rights compliance, and standards. (OSTP AI Bill of Rights; Executive Order on AI.)
• –NIST issued the AI Risk Management Framework (AI RMF 1.0) on 26 January 2023, with a Generative AI Profile released on 26 July 2024, providing a voluntary, rights‑preserving framework for mapping, measuring, and managing AI risks across the lifecycle. (NIST AI RMF 1.0; NIST RMF overview and GenAI profile.)
• –ISO/IEC standards codify management and risk practices, including ISO/IEC 42001:2023 (AI management systems) and ISO/IEC 23894:2023 (AI risk management guidance), complementing regulatory requirements and organizational governance. (ISO AI standards overview; ISO/IEC 23894:2023.)

Core principles and recurring issues

• –Fairness and non‑discrimination are central, but different formal criteria can be mutually incompatible in real‑world settings; seminal work shows inherent trade‑offs among commonly proposed risk‑score fairness conditions, guiding transparent policy choices. (Kleinberg, Mullainathan, Raghavan 2016.)
• –Accountability and redress require traceability, audits, and impact assessments across the AI lifecycle; professional bodies such as the IEEE and ACM have articulated principles for algorithmic transparency and accountability to operationalize oversight. (ACM 2017 principles; ACM 2022 update.)
• –Transparency and explainability practices include local and global interpretability (e.g., LIME; SHAP) and documentation norms (e.g., model cards; datasheets), improving scrutiny and communication of model behavior and limitations. (Ribeiro et al., LIME; Lundberg & Lee, SHAP; Model Cards, FAT* 2019; Datasheets for Datasets, CACM 2021.)
• –Privacy and data protection involve both regulatory safeguards and technical measures such as differential privacy and federated learning; data rights in the General Data Protection Regulation include limits on certain solely automated decisions with significant effects and rights to meaningful information about the logic involved. (Dwork & Roth 2014; McMahan et al. 2016; GDPR Article 22 on automated decisions and related case law (CJEU 2023–2024 excerpts).)
• –Safety, robustness, and security span development and deployment, including adversarial threats and malicious use; a multi‑stakeholder 2018 report mapped risks across digital, physical, and political domains and proposed mitigations. (The Malicious Use of AI.)

Sectoral contexts

• –Health: WHO guidance urges rigorous evaluation, transparency, inclusiveness, and accountability for AI in care delivery and public health; regulators have issued principles for machine‑learning medical devices. (WHO 2021 guidance; WHO 2023 LLM caution; FDA/Health Canada/MHRA GMLP.)
• –Justice and public sector: Studies and public investigations into risk assessment tools (e.g., COMPAS) catalyzed debate on statistical definitions of fairness, error disparities, and transparency requirements, illustrating how ethical design and governance interact with systemic inequities. (ProPublica “Machine Bias”; Kleinberg et al..)
• –Defense and security: Debates at the United Nations (CCW/GGE) and General Assembly address lethal autonomous weapons systems, emphasizing the need for meaningful human control and accountability in the use of force. (UN News, 2019 SG remarks; UNGA First Committee 2023 resolution coverage.)

Implementation tools and standards

• –Risk management and governance: The NIST AI RMF (2023) provides functions and profiles (including a 2024 Generative AI profile) to help organizations operationalize trustworthy AI, aligning with sectoral laws and international standards. (NIST AI RMF 1.0; NIST RMF overview.)
• –Management systems and audits: ISO/IEC 42001:2023 establishes AI management system requirements to integrate policy, roles, controls, and continuous improvement; ISO/IEC 23894:2023 gives guidance for AI‑specific risk management; national or industry programs (e.g., IEEE initiatives) address ethics‑by‑design and conformance. (ISO AI standards overview; ISO/IEC 23894; IEEE ethics resources.)
• –Documentation and assessment: Model cards and datasheets aid reproducibility and accountability; algorithmic impact assessments are promoted in governmental guidance and the EU AI Act’s risk‑based regime, which mandates obligations scaled to risk (including prohibitions and high‑risk controls). (Model Cards; Datasheets; EU AI Act, obligations and staged application.)

Environmental and social considerations

• –Training and serving modern models can be energy‑intensive; empirical analyses have quantified carbon and cost footprints in NLP research and urged efficiency improvements, while projections indicate rising electricity demand from data centers, partly driven by AI, underscoring sustainability considerations in ethical deployment. (Strubell et al., ACL 2019; Reuters on EPRI projections, 2024.)

Conceptual foundations and ongoing debates

• –Scholarship and civic literature debate long‑term safety, alignment, and control of advanced AI systems and near‑term social harms, including labor impacts, surveillance, and disinformation; these perspectives motivate plural ethical frameworks and policy instruments. ([Superintelligence](book://Nick Bostrom|Superintelligence: Paths, Dangers, Strategies|Oxford University Press|2014); [Weapons of Math Destruction](book://Cathy O'Neil|Weapons of Math Destruction|Crown|2016).)

Legal rights and safeguards

• –Data protection laws address automated decision‑making, meaningful information about logic, and rights to contest decisions; GDPR Article 22 and related jurisprudence clarify safeguards and human oversight requirements in consequential automated processing. (GDPR consolidated text, Art. 22; CJEU interpretations.)

Operational techniques and practices

• –Explainability, documentation, privacy‑enhancing technologies (e.g., differential privacy; federated learning), robust evaluation, red‑teaming, and post‑deployment monitoring are widely recommended across frameworks to make systems safer, fairer, and more accountable, supported by standards and regulatory guidance. (LIME; SHAP; Dwork & Roth 2014; McMahan et al.; Executive Order on AI.)