A taxonomy of Oracle AIs

Abstract

Oracle AIs are a popular concept in science fiction, but the term can be ambiguous. Some proposals for Oracle AIs describe systems that are inherently safe because they have no goals, while others propose systems that are safe only if their goal is perfectly aligned with human values. This article argues that both types of Oracle AI are either unsafe or FAI-complete (i.e., as difficult to create safely as a Friendly AI). It analyzes various proposals for Oracle AIs, including Advisors, Question-Answerers, and Predictors, and argues that many of these proposals are either unsafe or FAI-complete. – AI-generated abstract.