The Shutdown Problem: An AI Engineering Puzzle for Decision Theorists
Elliott Thornley (Global Priorities Institute, University of Oxford)
GPI Working Paper No. 10-2024, forthcoming in Philosophical Studies
I explain and motivate the shutdown problem: the problem of designing artificial agents that (1) shut down when a shutdown button is pressed, (2) don’t try to prevent or cause the pressing of the shutdown button, and (3) otherwise pursue goals competently. I prove three theorems that make the difficulty precise. These theorems suggest that agents satisfying some innocuous-seeming conditions will often try to prevent or cause the pressing of the shutdown button, even in cases where it’s costly to do so. I end by noting that these theorems can guide our search for solutions to the problem.
Other working papers
Crying wolf: Warning about societal risks can be reputationally risky – Lucius Caviola (Global Priorities Institute, University of Oxford) et al.
Society relies on expert warnings about large-scale risks like pandemics and natural disasters. Across ten studies (N = 5,342), we demonstrate people’s reluctance to warn about unlikely but large-scale risks because they are concerned about being blamed for being wrong. In particular, warners anticipate that if the risk doesn’t occur, they will be perceived as overly alarmist and responsible for wasting societal resources. This phenomenon appears in the context of natural, technological, and financial risks…
The paralysis argument – William MacAskill, Andreas Mogensen (Global Priorities Institute, Oxford University)
Given plausible assumptions about the long-run impact of our everyday actions, we show that standard non-consequentialist constraints on doing harm entail that we should try to do as little as possible in our lives. We call this the Paralysis Argument. After laying out the argument, we consider and respond to…
Will AI Avoid Exploitation? – Adam Bales (Global Priorities Institute, University of Oxford)
A simple argument suggests that we can fruitfully model advanced AI systems using expected utility theory. According to this argument, an agent will need to act as if maximising expected utility if they’re to avoid exploitation. Insofar as we should expect advanced AI to avoid exploitation, it follows that we should expected advanced AI to act as if maximising expected utility. I spell out this argument more carefully and demonstrate that it fails, but show that the manner of its failure is instructive…