Does AI become independent on its own?

A language model is a program that waits for an input. Without input, nothing happens. What is new are so-called AI agents. They break a task down into several steps, call on tools such as a web search or a calendar, and check their own result. From the outside, that looks independent.

This independence is borrowed. It reaches exactly as far as the access that a person has granted. An agent that may only read emails cannot trigger a bank transfer. An agent with access to a bank account can. Risks therefore arise mainly from rights that are granted too generously and from a lack of oversight.

Researchers deliberately test models in constructed extreme situations and sometimes observe unwanted behavior, for example when a model is told to achieve a given goal at any cost. Such tests show why limits are needed. Whoever uses AI decides what it is allowed to do, and checks what it actually does.