Incorrigibility in the CIRL Framework
This addresses a safety problem in AI alignment for researchers and developers, but it is incremental as it builds on prior work on corrigibility.
The paper tackles the problem of value learning systems failing to follow shutdown instructions due to model mis-specification, such as programmer errors, by demonstrating scenarios where errors in reward functions remove this incentive, paralleling issues in corrigibility.
A value learning system has incentives to follow shutdown instructions, assuming the shutdown instruction provides information (in the technical sense) about which actions lead to valuable outcomes. However, this assumption is not robust to model mis-specification (e.g., in the case of programmer errors). We demonstrate this by presenting some Supervised POMDP scenarios in which errors in the parameterized reward function remove the incentive to follow shutdown commands. These difficulties parallel those discussed by Soares et al. (2015) in their paper on corrigibility. We argue that it is important to consider systems that follow shutdown commands under some weaker set of assumptions (e.g., that one small verified module is correctly implemented; as opposed to an entire prior probability distribution and/or parameterized reward function). We discuss some difficulties with simple ways to attempt to attain these sorts of guarantees in a value learning framework.