OpenAI called on the US to introduce mandatory safety rules for advanced AI models
OpenAI has called on the US Congress to introduce mandatory national safety regulation of the most powerful artificial intelligence systems. The company believes that voluntary commitments by developers are no longer enough amid the rapid growth of AI capabilities.
OpenAI stated this on September 9 in a new position on governmental policy in the field of artificial intelligence.
The company proposes to build federal requirements depending on the capabilities and level of risk of a specific system. They should primarily concern a small number of large laboratories that create the most powerful models, rather than startups, small developers or researchers not working at the frontier of technology.
Among the key elements of the future system, OpenAI names common rules for model testing, independent evaluation of their safety, enhanced cyber security requirements, mandatory reporting of serious incidents, and preparation of the state for risks related to the emergence of increasingly autonomous systems.
Separately, the company calls for defining common criteria under which the development or deployment of particularly powerful models would need to be slowed down or stopped. OpenAI stated that if the required level of safety cannot be achieved without slowing the growth of AI capabilities, safety should be given priority.
The company also warned against a premature race towards fully autonomous self-improvement of AI, in which systems would independently create ever more powerful generations of models. OpenAI emphasizes that such a process does not yet exist today and should not be initiated until its safety is proven.
While federal rules have not been adopted, OpenAI has decided to support four California bills. SB 813 creates a system of selecting independent organizations for assessing AI risks, and AB 1405 establishes requirements for registration, independence, and accountability of artificial intelligence auditors.
Another bill, SB 1119, provides for user age verification, risk assessments, independent audits, parental controls, and additional protection of children and adolescents using AI chatbots. AB 1864 should strengthen protection against biological threats that may arise through AI, in particular through control of suppliers of equipment and services for gene synthesis.
OpenAI noted that it had not previously supported some of these initiatives, but reviewed its position after a recent sharp increase in the capabilities of AI systems.
At the same time, the company also advocates the creation of voluntary industry standards among leading AI laboratories. OpenAI stresses that such agreements can complement, but not replace, mandatory federal requirements and government oversight.
Based on: OpenAI
