Shutdown Safety Valves for Advanced AI
Artificial Intelligence
2026-03-10 v1 Machine Learning
Abstract
One common concern about advanced artificial intelligence is that it will prevent us from turning it off, as that would interfere with pursuing its goals. In this paper, we discuss an unorthodox proposal for addressing this concern: give the AI a (primary) goal of being turned off (see also papers by Martin et al., and by Goldstein and Robinson). We also discuss whether and under what conditions this would be a good idea.
Keywords
Cite
@article{arxiv.2603.07315,
title = {Shutdown Safety Valves for Advanced AI},
author = {Vincent Conitzer},
journal= {arXiv preprint arXiv:2603.07315},
year = {2026}
}