Schuett, Dreksler, Anderljung, McCaffary, Heim, Bluemke, Garfinkel
Towards best practices in AGI safety and governance
Schuett, Dreksler, Anderljung, McCaffary, Heim, Bluemke, Garfinkel. 2023. (View Paper → )
Many leading AI companies, including OpenAI, Google DeepMind, and Anthropic, aim to build artificial general intelligence (AGI) — AI systems that match or surpass human performance across various cognitive tasks. This pursuit could lead to the development and deployment of AI systems that pose significant risks. While some measures have been taken to mitigate these risks, best practices have yet to be established. To aid in identifying best practices, we sent a survey to 92 leading experts from AGI labs, academia, and civil society and received 51 responses. Participants were asked to rate their agreement with 50 statements about what AGI labs should do. The main finding is that participants, on average, agreed with all statements. Many statements received extremely high levels of agreement. For instance, 98% of respondents somewhat or strongly agreed that AGI labs should conduct pre-deployment risk assessments, dangerous capabilities evaluations, third-party model audits, safety restrictions on model usage, and red teaming. Ultimately, our list of statements may serve as a useful basis for efforts to develop best practices, standards, and regulations for AGI labs.
Much of AGI safety should remain the responsibility of the AGI labs. The 100 responsible AI practices listed in this paper are a helpful checklist for them to consider.
