Current approaches to building general-purpose AI systems tend to produce systems with both beneficial and harmful capabilities. Further progress in AI development could lead to capabilities that pose extreme risks, such as offensive cyber capabilities or strong manipulation skills. We explain why model evaluation is critical for addressing extreme risks. Developers must be able to identify dangerous capabilities (through "dangerous capability evaluations") and the propensity of models to apply their capabilities for harm (through "alignment evaluations"). These evaluations will become critical for keeping policymakers and other stakeholders informed, and for making responsible decisions about model training, deployment, and security.
@article{arxiv.2305.15324,
title = {Model evaluation for extreme risks},
author = {Toby Shevlane and Sebastian Farquhar and Ben Garfinkel and Mary Phuong and Jess Whittlestone and Jade Leung and Daniel Kokotajlo and Nahema Marchal and Markus Anderljung and Noam Kolt and Lewis Ho and Divya Siddarth and Shahar Avin and Will Hawkins and Been Kim and Iason Gabriel and Vijay Bolina and Jack Clark and Yoshua Bengio and Paul Christiano and Allan Dafoe},
journal= {arXiv preprint arXiv:2305.15324},
year = {2023}
}