The research and safeguards that aim to keep AI from causing harm.
AI safety covers the research and safeguards meant to stop AI from being misused or behaving in harmful, unexpected ways. It includes teaching models to refuse dangerous requests, attacking them on purpose before release to find weak spots, and keeping an eye on how they're used.
Experts disagree about which risks deserve the most attention. Some focus on problems happening now, like false information and unfair bias; others worry more about whether people can stay in control of much more powerful future systems. Countries have also started meeting about it, for example at the 2023 AI Safety Summit in the UK.
Think of it like this
Think of a roller coaster. You trust it because of many layers: lap bars, height limits and test runs every morning. AI safety works the same way: not one magic fix, but many layers stacked together.
For example
Before release, experts often try to trick a model into harmful answers on purpose. This is called red-teaming.
Common mix-ups
Mix-upAI safety just needs one gadget that makes the AI refuse dangerous requests.
ActuallyThe 2026 International AI Safety Report, backed by more than 30 countries and international organisations and written with more than 100 experts, concluded that skilled attackers can get around today's defences and that it's still unclear how well many safeguards actually work.
Mix-upExperts all agree on what the biggest danger is.
ActuallySome experts worry most about harms happening now, such as fake news or unfair treatment, while others worry most that people may find AI harder to control in future. Because views differ, places to check things together are appearing: at the 2024 Seoul summit, co-hosted by Korea and the UK, countries agreed to link their AI safety institutes into one network.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- Like the safety features on a roller coaster, what layers does AI need?
- Activity
- Each group lists roller-coaster safety features (lap bar, height limit, test runs…) and, using this entry, draws a line from each to a matching AI safety layer. Then open the Bletchley (E0284) and Seoul (E0374) declaration records and write one line on what the countries agreed.
Records about this
Source-checked moments from AI history.
Brings the format of an IPCC-style international consensus document to the AI risk debate. A shared factual basis for regulatory debates in each country
An international AI governance milestone co-hosted by Korea
The starting point of the line of international AI safety summits, and the source of the mandate for the International AI Safety Report