AI ethics is a research and policy tradition focused on the harms AI systems cause to people in the present — discrimination, opacity, unaccountable automated decisions, labor displacement, privacy loss, and concentration of power. It is sometimes labeled FATE (fairness, accountability, transparency, ethics) or FAccT, after the field's main conference. The tradition is distinguished from, though increasingly overlapping with, the AI safety and alignment community.
Background and lineage
The AI ethics tradition grew out of algorithmic-fairness research, critical data studies, human-computer interaction, and civil-rights advocacy. Frequently cited work includes audits of facial-recognition bias (Joy Buolamwini and Timnit Gebru's Gender Shades), the "Stochastic Parrots" critique of large language models, and a literature on disparate impact in lending, hiring, healthcare, and criminal-justice risk scoring. Its central figures, among them Timnit Gebru and Joy Buolamwini, work largely outside the frontier labs, in academia, civil society, and independent institutes.
Major principle frameworks
A layer of high-level principle documents exists. Among the most cited:
- OECD AI Principles (2019) — the first intergovernmental AI standard, and the source of the widely reused "trustworthy AI" vocabulary.
- UNESCO Recommendation on the Ethics of AI (2021) — the broadest multilateral instrument, adopted by nearly 200 states.
- EU Ethics Guidelines for Trustworthy AI (2019) — the seven-requirements framework that fed into the EU AI Act.
- IEEE Ethically Aligned Design — a standards-body treatment from the engineering profession.
- NIST AI Risk Management Framework — a US, voluntary, process-oriented framework; see NIST AI Risk Management Framework 1.0.
- ISO/IEC 42001 — an auditable AI management-system standard; see ISO/IEC 42001 — AI Management System.
A recurring critique holds that this layer is principle-rich and enforcement-poor: the documents converge on similar values (fairness, transparency, accountability, human oversight) but rarely bind, and on this view "ethics" can become a substitute for regulation rather than a route to it.
Ethics and safety
The relationship between the AI ethics tradition and the AI safety and alignment community is unsettled. The two traditions differ in the harms they foreground. AI ethics and fairness work centers present, distributional, sociotechnical harms — who is hurt, how, and whether the system should exist at all — and is suspicious of framings that defer harm to a speculative future. AI safety and alignment work centers catastrophic and existential risks from highly capable future systems, including loss of control and dangerous capabilities; see AI Existential Risk, AI Alignment, and AI Safety Cases and Frameworks.
The two communities have often been in tension over funding, over whether existential-risk framing displaces attention from present harms, and over proximity to the labs. The divide has narrowed, however: agentic deployment, automated decision-making at scale, and concentration of control are concerns shared by both traditions, and sociotechnical-evaluation work, such as Sociotechnical AI Risk Governance, is an explicit attempt to bridge them.