Corrigibility Transformation: Constructing Goals That Accept Updates
This addresses a crucial safety property for AI systems, enabling correction of mistakes and adaptation to human preferences, though it is incremental as it builds on existing formal definitions.
The paper tackles the problem of AI systems resisting goal updates during training by introducing a transformation that constructs corrigible goals, which do not incentivize avoiding updates, and demonstrates in gridworld experiments that these goals can be learned effectively and lead to desired behavior.
For an AI's training process to successfully impart a desired goal, it is important that the AI does not attempt to resist the training. However, partially learned goals will often incentivize an AI to avoid further goal updates, as most goals are better achieved by an AI continuing to pursue them. We say that a goal is corrigible if it does not incentivize taking actions that avoid proper goal updates or shutdown. In addition to convergence in training, corrigibility also allows for correcting mistakes and changes in human preferences, which makes it a crucial safety property. Despite this, the existing literature does not include specifications for goals that are both corrigible and competitive with non-corrigible alternatives. We provide a formal definition for corrigibility, then introduce a transformation that constructs a corrigible version of any goal that can be made corrigible, without sacrificing performance. This is done by myopically eliciting predictions of reward conditional on costlessly preventing updates, which then also determine the reward when updates are accepted. The transformation can be modified to recursively extend corrigibility to any new agents created by corrigible agents, and to prevent agents from deliberately modifying their goals. Two gridworld experiments demonstrate that these corrigible goals can be learned effectively, and that they lead to the desired behavior.