OpenAI Practices for Governing Agentic AI Systems
TL;DR
OpenAI has introduced a framework of practices for governing agentic AI systems to ensure they are integrated responsibly into society. This framework aims to establish baseline responsibilities and safety best practices for all parties involved in the agentic AI life-cycle to mitigate risks of harm while enabling the efficiency and effectiveness of complex goal pursuit with limited supervision.
Defining Agentic AI Systems
Agentic AI systems are defined as AI systems capable of pursuing complex goals with limited direct supervision. These systems possess the potential to help users achieve their goals more efficiently and effectively, but their autonomy increases the risk of potential harm.
Governance Framework and Life-Cycle Responsibilities
To manage the risks associated with agentic AI, OpenAI suggests that governance must focus on the defining the roles and responsibilities of the parties involved in the agentic AI system life-cycle. This includes agreeing on a set of baseline responsibilities and safety best practices for each party to ensure operational safety and accountability.
Operationalizing Safety Practices
OpenAI identifies that while initial practices for keeping agents' operations safe and accountable are offered, there are significant uncertainties and questions surrounding the operationalization of these practices. These must be addressed before such practices can be codified into formal standards.
Indirect Impacts of Wide-Scale Adoption
Beyond the immediate operational safety of individual agents, the wide-scale adoption of agentic AI systems is expected to create indirect impacts. OpenAI notes that these impacts will likely necessitate the development of additional, broader governance frameworks to manage the systemic effects of autonomous AI agents in society.
Research Grant Program
OpenAI launched a research grant program awarding between $10,000 and $100,000 to fund research into the impacts of agentic AI systems and the practices required to make them safe. (Note: This grant program is now closed).