Foundation Model Convening

The rapid evolution of artificial intelligence (AI) technology offers immense opportunity to advance human welfare but also poses novel threats. Foundation models (FMs), unlike other AI, are trained on large datasets that show competence across a wide variety of domains and tasks, such as generating text, audio, video, and images. The generalized competence of FMs is the root of both their potential benefits and their potential risks. In terms of potential positive outcomes, FMs could provide benefits across a wide variety of sectors, such as education. In terms of potential negative outcomes, FMs could allow the creation of chemical and biological weapons or amplify disinformation campaigns that undermine democratic elections.

RAND and the Carnegie Endowment for International Peace hosted a convening on FMs on November 10–12, 2023. The convening covered such topics as model cybersecurity, developer ethical norms, international governance, legal liability, threat assessment, and reporting and other forms of information-sharing. These conference proceedings capture the industry, academic, and think-tank perspectives emerging from the workshops to inform government, civil society, industry, and the broader public discussion about artificial intelligence safety and security.