Glossary
AI Safety
Work aimed at preventing AI systems from causing severe harm, including accidents, misuse and loss of meaningful human control.
Last reviewed 2026-08-22
In plain language
AI safety covers a spectrum: conventional software and product safety; robustness and evaluation of machine-learning systems; biosecurity and cybersecurity interfaces; and longer-term debates about highly capable general systems. Communities disagree about timescales, threat models and whether safety should be framed primarily as a technical, corporate or public-policy problem. Safety is related to, but not identical with, ethics, security or human-rights protection.
Why it matters for AI governance
Safety arguments now shape regulation, research funding and international diplomacy. Policymakers need to specify which harms, which systems and which evidence standards they mean.
Authoritative sources
- Hiroshima AI Process (opens in a new tab) — G7 / Government of Japan