Imperfect Recall and AI Delegation
Eric Olav Chen (Global Priorities Institute, University of Oxford), Alexis Ghersengorin (Global Priorities Institute, University of Oxford) and Sami Petersen (Department of Economics, University of Oxford)
GPI Working Paper No. 30-2024
A principal wants to deploy an artificial intelligence (AI) system to perform some task. But the AI may be misaligned and pursue a conflicting objective. The principal cannot restrict its options or deliver punishments. Instead, the principal can (i) simulate the task in a testing environment and (ii) impose imperfect recall on the AI, obscuring whether the task being performed is real or part of a test. By committing to a testing mechanism, the principal can screen the misaligned AI during testing and discipline its behaviour in deployment. Increasing the number of tests allows the principal to screen or discipline arbitrarily well. The screening effect is preserved even if the principal cannot commit or if the agent observes information partially revealing the nature of the task. Without commitment, imperfect recall is necessary for testing to be helpful.
Other working papers
Longtermism, aggregation, and catastrophic risk – Emma J. Curran (University of Cambridge)
Advocates of longtermism point out that interventions which focus on improving the prospects of people in the very far future will, in expectation, bring about a significant amount of good. Indeed, in expectation, such long-term interventions bring about far more good than their short-term counterparts. As such, longtermists claim we have compelling moral reason to prefer long-term interventions. …
Towards shutdownable agents via stochastic choice – Elliott Thornley (Global Priorities Institute, University of Oxford), Alexander Roman (New College of Florida), Christos Ziakas (Independent), Leyton Ho (Brown University), and Louis Thomson (University of Oxford)
Some worry that advanced artificial agents may resist being shut down. The Incomplete Preferences Proposal (IPP) is an idea for ensuring that doesn’t happen. A key part of the IPP is using a novel ‘Discounted REward for Same-Length Trajectories (DREST)’ reward function to train agents to (1) pursue goals effectively conditional on each trajectory-length (be ‘USEFUL’), and (2) choose stochastically between different trajectory-lengths (be ‘NEUTRAL’ about trajectory-lengths). In this paper, we propose evaluation metrics…
Cassandra’s Curse: A second tragedy of the commons – Philippe Colo (ETH Zurich)
This paper studies why scientific forecasts regarding exceptional or rare events generally fail to trigger adequate public response. I consider a game of contribution to a public bad. Prior to the game, I assume contributors receive non-verifiable expert advice regarding uncertain damages. In addition, I assume that the expert cares only about social welfare. Under mild assumptions, I show that no information transmission can happen at equilibrium when the number of contributors…