AI Safety

What AI Safety is

AI safety covers present-day harms such as misinformation, bias and misuse, and longer-horizon questions about oversight of increasingly capable systems. Both agendas share techniques: evaluation, interpretability, guardrails and human oversight.

How it works

Work includes capability and dangerous-capability evaluations before release, alignment training, interpretability research, staged deployment, monitoring and incident response.

Why it matters

Safety practice determines whether AI deployment in medicine, finance, infrastructure and education is defensible, and it increasingly carries legal weight.

Common uses

• →Pre-deployment evaluation
• →Deployment policy and staged rollout
• →Model cards and system documentation