What Is AI Safety?

Safety concerns harmful outcomes, whether caused by error, misuse or unsuitable deployment. Security concerns protection against unauthorised access, manipulation and attacks. Alignment concerns whether a system's behaviour reflects intended goals and constraints.

These areas overlap. A secure system may still make a harmful mistake; a well-behaved model may still be vulnerable to an attack. Work in one area does not remove the need for the others.

Reliability means consistent, dependable performance under the conditions in which a system is used. A tool that succeeds most of the time can still be unsuitable for a high-stakes task if its failures are hard to detect.

A language model may produce an unsupported medical claim or invent a legal citation. This kind of unsupported output is often called a hallucination. Useful safeguards can include source checking, uncertainty handling, constrained workflows and qualified human review. Whether they are sufficient depends on the use case.

Robustness concerns performance under perturbations, unfamiliar inputs or changing conditions. A system that works on curated examples may fail when the input is incomplete, ambiguous or adversarial.

Testing should include realistic failure cases, not just average benchmark performance. Operators also need ways to detect unexpected behaviour and stop or limit the system when it leaves the conditions for which it was evaluated.

AI can be used to automate deception, create malicious content or assist harmful activity. Some capabilities may be dangerous because they lower the effort or expertise required for an attack.

Evaluation asks what a system actually enables, including when it has tools and access to external resources. A refusal in a simple test is not a complete defence against misuse, just as unrestricted access is not a neutral deployment choice.

Safety evaluations can combine task tests, red teaming, expert review and analysis of incidents. Dangerous-capability evaluations focus on abilities that could enable serious harm, while other assessments examine everyday reliability and bias.

No test suite covers every possible situation. Evaluations need clear scope, appropriate threat models and follow-up as systems or operating conditions change. Independent scrutiny can help reveal blind spots.

Deployment decisions may include access limits, tool permissions, monitoring, rollback options and documentation of known weaknesses. These are operational controls, not proof that a model is intrinsically safe.

More capable systems might create new risks through greater autonomy, influence or access to important infrastructure. Discussions of superintelligence include uncertain scenarios involving loss of control or severe, widespread harm.

Those possibilities should be assessed with explicit assumptions, not presented as established outcomes. Addressing future risks need not displace work on existing harms; both require evidence, proportionate safeguards and accountability.

Safety is an ongoing process of learning where a system fails and reducing the consequences. It is not a label that can be permanently attached after one successful test.