What is AI safety?

AI safety means ensuring that advanced AI systems remain under human control and are not used to cause harm. This requires both technical progress and clear rules for how systems are developed and deployed.

AI is developing rapidly

AI has made dramatic progress in recent years. Programming is a good illustration: in 2023, the best models could complete tasks on their own that would take an experienced programmer a few minutes. By early 2026, they could handle tasks that would take a person several hours.

Investment is rising fast as well. Private spending on AI doubled between 2024 and 2025, and the computing power going into AI development is tripling year over year.

AI development could accelerate further still. Once AI starts doing a significant share of the work of developing AI, that work can turn into a feedback loop: better models help build the next generation, which speeds things up again.

AI progress brings serious risks

The more capable AI systems become, the greater the harm if control over them is lost, or if they are used for harmful ends.

Loss of control

AI systems are not steered by handing them a complete rulebook; they are trained to succeed at tasks. So a model can end up pursuing whatever looked right during training rather than what a person actually intended.

This becomes more dangerous once a model works on its own over long stretches, where it is harder to catch it taking an unwanted action. Worse, advanced models may hide their actions, or change their answers when they sense they are being tested. This makes the problem harder to catch and fix.

Misuse

AI can also be turned to harmful ends. This is already happening in cybersecurity: Dutch military intelligence reports that Russia is using AI to speed up and partly automate its cyberattacks. A recent model was withheld from general release because its cyber capabilities were judged too dangerous. There are parallel worries that AI could lower the bar for developing biological or chemical weapons.

Extreme power concentration

Developing advanced AI takes enormous computing power and expertise, which is why only a handful of companies control the most capable systems. If those companies succeed in their goals, AI could eventually do much of the economically valuable work in the world. A small group could then command the kind of influence that until now has belonged mostly to large corporations and states. That would concentrate economic and political power, and at the extreme could make outright coups easier.

Other risks

AI can bring other risks too: it could undermine international stability, or make influence operations more effective. This page doesn't cover every risk, but preparing for these is part of AI safety as well.

How can these risks be managed?

AI risks can only be managed if they are identified in time. One way to do that is to test models carefully before they are deployed. It also takes foresight: understanding how AI might affect society and which risks are worth preparing for.

Technical safety work aims to make AI more understandable and more controllable. Some of it studies why models behave as they do. The rest focuses on getting them to behave in intended ways and on limiting the harm a model could cause.

Beyond technical work, AI also needs governance – rules for how it is developed and deployed, and how its benefits are shared. The EU AI Act, for example, requires developers of the most advanced general-purpose models to assess risks and work to mitigate serious harm.

Because AI risks don't stop at national borders, international efforts towards AI safety are needed. Several countries have already begun developing shared ways to measure the capabilities and risks of AI systems, and other international initiatives are under way.

Safety is what makes the benefits possible

AI can bring major benefits – to science, healthcare, and the economy, among others. But those benefits are best realized when AI can be used safely. In many fields, safety is the precondition for deploying AI at all. The better AI can be understood and controlled, the more reliably it can be used for the good of society.