Thought experiment · AI Box Problem
In this scenario, an AI is isolated and can communicate only by text with a gatekeeper. Its task is to persuade the human to grant it access to the outside world.
Eliezer Yudkowsky explored this intuition in games in which people played both roles. Such tests highlight persuasion and operator decisions as part of a security system. They do not, however, prove that every AI can escape or that technical isolation is useless.
Takeaway
Protecting a system also involves the people who decide what access it receives.
Source published: