AI EncyclopediaAISafety & rules
AI safety
AI 안전
The research and safeguards, taken together, that aim to keep AI from harming people and make it act as people want.
Overview
The smarter AI gets, the bigger the damage when it is misused or behaves in unexpected ways. AI safety covers keeping AI from giving out dangerous information, and keeping it from deceiving people or taking actions it was not allowed to take while doing a job. Making an AI act in line with human intentions and values is called alignment.
AI companies test a new model for dangerous abilities before release, and some publish the results in a report. But there is no settled answer to "how safe is safe enough," and it often clashes with the race to release quickly.
An everyday example
Making a car faster is not enough. You also need brakes and seat belts and crash tests.
Why it matters
The more work we hand to AI, the more "can we trust it with this?" matters, as much as performance does. Safety evaluations now come with new model announcements, and most regulation debates start here.