Legislation has not reached "self-control" first: AI giants will promise product safety, OpenAI supports licensing and approval of AI system development

Author: Li Dan

Image source: Generated by Unbounded AI tool

Although the United States has not yet introduced specific legislation, under the pressure of the Biden administration, some giants and rising stars in the field of artificial intelligence (AI) OpenAI have taken the initiative to propose "self-control."

On Thursday, July 20th, Eastern Time, the media quoted insiders as saying that at the request of the White House, some companies, including Microsoft, Google’s parent company Alphabet, and OpenAI, will make public commitments this Friday to ensure their safe use of AI technology and develop and deploy AI responsibly.

The draft White House document shows that the companies' pledges will remain in effect until the U.S. Congress passes legislation to address AI-related issues. These promises focus on generative AI and the most powerful existing AI models, and even possible future models. In the draft, the White House proposed eight commitment points:

• Allow independent experts to act as imaginary enemies, try to make AI models produce bad behavior, carry out so-called red-teaming against blue and blue, and improve defense capabilities.
• Share trust and security information with governments and other companies.
• Use watermarks on audio and video content to help identify AI-generated content.
• Invest in cybersecurity measures.
• Encourage third parties to discover security holes.
• Report social risks such as misuse and bias.
• Prioritize research on the social risks of AI.
• Use the most advanced AI system, the so-called cutting-edge model, to solve the most serious social problems.

The media said that before this Friday, the specific content of the commitment may still have changes. But the promises are expected to be in line with those made at a meeting of some of the top AI companies in May.

Also on Thursday, another report said that an internal memo showed that OpenAI supports the implementation of a license system, that is, to develop advanced AI systems, developers need to first obtain a license issued by the government. OpenAI also hinted at a willingness to be more open about the data used to train AI imagers such as DALL-E, saying it is committed to incorporating an origin-based approach by the end of this year.

However, Makanju, OpenAI's vice president of global affairs, recently said that the licensing regime supported by OpenAI is limited to AI models that are more powerful than GPT-4, because OpenAI wants to ensure that small startups do not experience too much regulatory pressure.

The comment stated that the voluntary commitment itself has shown that the Biden administration has limitations in guiding AI models to avoid abuse. The U.S. Congress will not be able to pass bills and legislation in the short term, and regulation in the field of AI cannot be delayed.

Wall Street has mentioned that in May, the White House convened a meeting with the CEOs of Alphabet, Microsoft, and OpenAI to discuss AI security issues. Biden reiterated in the invitation letter at the time that he hoped that these companies must ensure product safety before releasing products. He also said there were concerns that rapidly advancing AI technology could invade privacy and be biased, spreading rumors and disinformation.

Last month, Biden once again reiterated the need to ensure safety before the release of AI products, saying that the national security and economic risks brought about by AI need to be addressed. White House chief of staff Jeff Zients said that month that some leading companies in the AI field are making efforts to cooperate with the Biden administration and will soon announce a series of privacy and security commitments. He also said that the regulatory process may be relatively slow, but we cannot afford to wait a year or two.