AI Safety is the study of how to make sure increasingly capable AI systems do what we intend, remain understandable and controllable, and are governed well. This page is a guided entry point. Read the overview, pick the area closest to your background, and follow the recommended reading.

Two layers of one problem

People often use “AI Safety”, “responsible AI” and “AI governance” interchangeably. They overlap, but they ask different questions. Keeping them distinct helps you find the right literature and the right tools.

Technical AI Safety

Can we reliably control and understand increasingly capable AI systems — including systems more capable than the people supervising them?

  • Alignment and training methods
  • Interpretability
  • Evaluations and robustness
  • Agent control and oversight

Responsible AI and governance

Are deployed systems fair, accountable and lawful, and do institutions have the rules and capacity to keep them that way?

  • Fairness, transparency and rights
  • Risk management and standards
  • Regulation, liability and data protection
  • Institutional and international coordination

Security and misuse sit across both: a well-aligned model can still be attacked, and a well-governed organisation can still deploy a system that is easy to misuse.

Seven areas to know

Alignment and specification

Technical AI Safety

Training gives a model an objective, but objectives are proxies for what we actually want. Alignment research studies how systems come to pursue goals that differ from the intended ones — through reward hacking, poorly specified objectives or goals that generalise in unexpected ways — and develops training methods, such as learning from human feedback, that narrow the gap.

Interpretability

Technical AI Safety

Modern models are trained, not programmed, so even their developers cannot read off why they behave as they do. Mechanistic interpretability tries to reverse-engineer the internal computations of a network into human-understandable parts, so that problems can be detected from the inside rather than only from behaviour.

Evaluation and robustness

Technical AI Safety

Evaluations measure what a system can do and what it tends to do, including dangerous capabilities. Robustness research asks whether good behaviour holds up under distribution shift and deliberate attack. Together they underpin most decisions about whether a model is safe to deploy.

Scalable oversight and human agency

Technical AI Safety and human factors

As systems become more capable, checking their work gets harder. Scalable oversight develops ways for people — often assisted by other AI systems — to supervise tasks they cannot easily evaluate themselves. Human agency asks how to keep meaningful human understanding and choice in the loop.

Recommended first reading

Accessible starting points, chosen by our editors. No technical background needed.

Zoom In: An Introduction to Circuits

Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, Shan Carter, Distill, 2020

A highly visual introduction to the idea that neural networks can be understood in terms of features and the circuits that connect them, using examples from an image classifier.

ExplainerIntroductoryAbout 1 hour

International AI Safety Report 2025

Chaired by Yoshua Bengio, with an international expert advisory panel, UK Government (GOV.UK), 2025

A scientific assessment of the capabilities and risks of general-purpose AI and of the techniques available to manage them, written by an independent panel of experts nominated by many countries and international organisations.

Policy or legal textIntroductoryAbout 6 hours (summary: 30 minutes)

All introductory resources

Choose a learning pathway