Simeon Campos (SaferAI) and Romain Laurent (same) have posted “A Definition of General-Purpose AI Systems: Mitigating Risks from the Most Generally Capable Models” on SSRN. Here is the abstract:
The European Union (EU) is currently going through the legislative process on the EU AI Act – the first bill intended to regulate Artificial Intelligence (AI) comprehensively in a major jurisdiction. The bill includes provisions to manage risks of generally capable AIs classified as “General Purpose AI Systems” (GPAIS). We believe that this crucial aspect of the act could be improved by focusing the definition more on the most generally capable systems, which bring very specific risks. The Future of Life Institute (FLI) proposed a definition of GPAIS to better target these models, a significant step in the right direction. Expanding on FLI’s proposal, this paper introduces a new definition of GPAIS, which serves to clearly differentiate between narrow and general systems, and cannot be easily exploited by GPAIS providers who may wish to avoid new regulatory constraints.
This paper consists of two sections. The first section discusses the specific risks of GPAIS, including unpredictability, adaptability, and the potential for emergent capabilities. The second section presents the new definition of GPAIS, and explains the changes made and how they address the risks presented in the first section. The EU AI Act could set a global standard for AI-related risk management. The aim of this document is to help inform AI Act draft reviews and improve the ability to mitigate risks from the most generally capable models to protect stakeholders in the EU and globally.