New Method for Creating Corrigible AI Goals

A new arXiv paper (2510.15395v2) introduces a transformation that constructs corrigible versions of nearly any AI goal, ensuring the AI remains open to updates and shutdown without sacrificing performance. The method elicits predictions of reward conditional on costlessly preventing updates, and the target is pursued myopically. The authors demonstrate that these goals achieve optimal performance among corrigible goals, incentivize mid-action overrides, and disincentivize deliberate resistance to updates. This addresses a critical gap in AI safety literature, as previous work did not specify goals that are both corrigible and competitive.