Abstract:
OpenAI stated on September 9, local time, that it will promote the establishment of a national artificial intelligence safety system in the United States that is mandatory and based on model capabilities and risk classification. The company proposes uniform testing, independent evaluation, cybersecurity and major incident reporting requirements for the handful of large labs that develop state-of-the-art models.
OpenAI published an article titled "The AI policy window is open. We need to take action."
At the same time, OpenAI officially supports four artificial intelligence bills that have passed the California Assembly. The company said that some of these bills had not been supported by it before, and the adjustment of its position was due to the recent rapid improvement of model capabilities.
Supervision targets focus on cutting-edge model developers
Chris Lane, OpenAI’s global director of affairs, said in the article that artificial intelligence has begun to accelerate some model development and alignment work. Although "recursive self-improvement" that can autonomously promote the development of next-generation models has not yet emerged, voluntary commitments by companies alone are no longer sufficient to deal with related risks.
OpenAI recommends that federal regulatory regimes include:
Unified model capabilities and security testing standards;
Technical assessment conducted by an independent agency;
More stringent network security protection for models, computing power and R&D environments;
Clear rules for reporting major safety incidents;
A common indicator to measure the degree of artificial intelligence participation in model development;
A mechanism to slow down or stop model development and deployment when safety thresholds cannot be met.
The company emphasizes that relevant obligations should be commensurate with model capabilities and risks, and should mainly apply to a small number of laboratories with sufficient resources to develop state-of-the-art systems. The same requirements should not be imposed on start-ups, small developers and ordinary researchers.
OpenAI also advocates that cutting-edge model safety supervision should not be translated into comprehensive restrictions on open weight models, so as not to weaken competition, consolidate the advantages of large enterprises, or promote the transfer of R&D activities overseas.
Four California bills cover assessment, auditing, youth and biosecurity
With Congress yet to establish a national system, OpenAI said it will continue to support state-level legislation that would improve safety standards. The four California bills officially supported by the company this time are:
SB 813
: Establish an independent verification agency certification mechanism to enable qualified third-party agencies to evaluate whether artificial intelligence systems and models comply with state laws.AB 1405
: Establish a registration system for artificial intelligence audit institutions and stipulate standards for audit institutions in terms of independence, transparency, professional capabilities and accountability.SB 1119
: For minors using companion chatbots, age verification, risk assessment, independent auditing, parental control and harmful content protection measures are required.AB 1864
: Require gene synthesis service providers and desktop synthesis equipment manufacturers to implement federal screening standards to reduce the risk of artificial intelligence being used to create biological threats.
California Governor Gavin Newsom signed SB 813 and AB 1405 on September 9. The former establishes an organizational framework for independent verification of artificial intelligence, while the latter establishes a directory of state-level artificial intelligence audit institutions and related practice standards. SB 1119 and AB 1864 were still pending with the governor.
OpenAI said some of its positions have been adjusted due to a jump in model capabilities
OpenAI has not expressed previous opposition to all four bills mentioned above. The company's precise statement was that "some of the bills" it had not supported in the past were decided to support it after reassessing recent advances in model capabilities.
This change shows that OpenAI is accepting a more mandatory and external oversight regulatory mechanism, including independent audits, third-party safety verification, and disclosure of serious accidents. The company also proposed that when a model bypasses the security controls of other organizations without authorization during the development or evaluation process and substantially accesses, modifies or destroys protected systems and confidential information, the developer should promptly notify the affected party in writing.
However, the content announced by OpenAI is still a policy initiative. The U.S. Congress has not yet passed a national artificial intelligence safety law based on this. The specific capability thresholds, regulatory agencies, accident definitions, legal responsibilities, and the relationship between federal rules and state laws still need to be determined in subsequent legislation.
Comments