Games for AI Control: Models of Safety Evaluations of AI Deployment Protocols
Griffin, Charlie, Thomson, Louis, Shlegeris, Buck, Abate, Alessandro
–arXiv.org Artificial Intelligence
To evaluate the safety and usefulness of deployment protocols for untrusted AIs, AI Control uses a red-teaming exercise played between a protocol designer and an adversary. This paper introduces AI-Control Games, a formal decision-making model of the red-teaming exercise as a multi-objective, partially observable, stochastic game. We also introduce methods for finding optimal protocols in AI-Control Games, by reducing them to a set of zero-sum partially observable stochastic games. We apply our formalism to model, evaluate and synthesise protocols for deploying untrusted language models as programming assistants, focusing on Trusted Monitoring protocols, which use weaker language models and limited human assistance. Finally, we demonstrate the utility of our formalism by showcasing improvements over empirical studies in existing settings, evaluating protocols in new settings, and analysing how modelling assumptions affect the safety and usefulness of protocols.
arXiv.org Artificial Intelligence
Sep-12-2024
- Country:
- North America > United States
- New York > New York County
- New York City (0.14)
- Massachusetts > Middlesex County
- Cambridge (0.04)
- New York > New York County
- Europe
- United Kingdom > England
- Oxfordshire > Oxford (0.04)
- Cambridgeshire > Cambridge (0.04)
- Sweden > Stockholm
- Stockholm (0.04)
- United Kingdom > England
- North America > United States
- Genre:
- Research Report (1.00)
- Industry:
- Leisure & Entertainment > Games (0.67)
- Information Technology (0.45)
- Technology: