Thinking Inside the Box: Controlling and Using an Oracle AI

Thus far, the vast majority of all AI safety research has focused on the case of a single AI system, often (implicitly) in the context of a single human.

“Thus far, the vast majority of all AI safety research has focused on the case of a single AI system, often (implicitly) in the context of a single human (see, e.g., Amodei et al., 2016; Armstrong et al., 2012; Christiano et al., 2018; Dalrymple et al., 2024; Hadfield-Menell et al., 2016; Hendrycks et al., 2021; Leike et al., 2018).”

The GS AI approach remains agnostic about both how the core AI system was produced, and about what containment or boxing methods are used.

“Note that the GS AI approach remains agnostic about both how the”

We may require that the AI has no incentive to influence any part of the external world, or to find out any information about the external world.

“For example, we may require that the AI has no incentive to influence any part of the external world, or to find out any information about the external world (e.g.Armstrong et al., 2012;Armstrong & O'Rorke, 2018;Armstrong & O'Rourke, 2018;Everitt et al., 2021;van Merwijk et al., 2022).”