Anton Korinek's report (December 8, 2021) presents two control problems with advanced AI:
- Direct control, that an AI system does what its operator wants it to do
- Social control, that an AI system does what the rest of society wants it to do
He uses the Facebook Files to show examples of failures in both direct and social control.
To solve these problems, Korinek proposes a new federal agency to regulate advanced AI:
Throughout our history, whenever we developed new technologies that posed new hazards for society, our nation has made it a habit to establish new regulatory bodies and independent agencies endowed with world-class expertise to oversee and investigate the new technologies. For example, the National Transportation Safety Board (NTSB) and the Federal Aviation Administration (FAA) were established at the onset of the age of aviation; or the Nuclear Regulatory Commission (NRC) was established at the onset of the nuclear age. By many measures, advanced artificial intelligence has the potential to be an even more powerful technology that may impose new types of hazards on society, as exemplified by the Facebook Files.
Given the rise of artificial intelligence, it is now time to establish a federal agency to oversee advanced artificial intelligence—an AI Control Council that is explicitly designed to address the AI Control Problem, i.e. to ensure that the ever more powerful AI systems we are creating act in society’s interest. To be effective in meeting this objective, such a council would need to have the ability to (i) pursue solutions to the direct AI control problem and (ii) to oversee and when necessary regulate the way AI is used across the U.S. economy to address the social control problem, all while ensuring that it does not handicap advances in AI.