WASHINGTON >> The Biden administration takes a new front in efforts to protect U.S. AI from China and Russia with preliminary plans to install guardrails around cutting-edge AI models that are the core software of artificial intelligence systems like ChatGPT. They are ready to open up, sources said. .
The Commerce Department is considering tightening new regulations to limit the export of proprietary and closed-source AI models, where the software and data used for training are kept secret, three people familiar with the matter said. people revealed.
Any measure would complement a series of measures taken over the past two years to block the export of advanced AI chips to China, with the aim of slowing the Chinese government's development of cutting-edge technology for military purposes. Still, it will be difficult for regulators to keep up with the industry's rapid development.
The Commerce Department declined to comment, and the Russian embassy in Washington did not respond to a request for comment. The Chinese embassy called the move “a typical act of economic repression and unilateral bullying, which China firmly opposes,” adding that it would take “necessary measures” to protect its interests. .
Currently, major U.S. AI companies such as Microsoft-backed OpenAI, which has developed the most powerful closed-source AI models, Alphabet's Google DeepMind, and rival Anthropic, are making their models available to nearly everyone in the world without government oversight. There's nothing stopping you from selling.
Government and private sector researchers believe that U.S. adversaries can use this model, which mines vast amounts of text and images to summarize information and generate content, to launch offensive cyberattacks and create powerful There are concerns that it could be used to create biological weapons.
One of the sources said new export restrictions would likely target Russia, China, North Korea and Iran. Microsoft said in a February report that hacking groups affiliated with the Chinese government, the North Korean government, Russian military intelligence, and Iran's Revolutionary Guard Corps were using large-scale language models to perfect their hacking campaigns. He said he was tracking.
calculation ability
To develop export controls for AI models, the U.S. may rely on thresholds contained in an AI executive order issued last October based on the amount of computing power needed to train the models, officials said. said. Once that level is reached, the developer must report his AI model development plan and submit test results to the Department of Commerce.
That computational power threshold could be the basis for determining which AI models are subject to export controls, according to two U.S. officials and another person briefed on the discussions. They declined to give their names because details have not been made public.
According to EpochAI, a research organization that tracks AI trends, Google's Gemini Ultra is seen as meeting that standard, but is still considered below the mark, so even if it is used. , it is likely to only restrict the export of models that have not yet been released. .
Officials emphasized that the agency is far from finalizing the proposed rule. But the fact that such a move is being considered highlights gaps in the U.S. government's efforts to thwart Beijing's AI ambitions, despite the significant challenge of imposing a robust regulatory regime on rapidly evolving technology. It shows that you are trying to fill in the
As the Biden administration considers competition with China and the dangers of advanced AI, AI models are “obviously one of the tools and one of the potential challenges that needs to be considered here.” said Peter Harrell, former National Security Council member. Formal. He added: “In fact, it remains to be seen whether, practically speaking, we can turn it into an export controllable choke point.”
Biological weapons and cyber attacks?
U.S. intelligence agencies, think tanks, and academics are increasingly concerned about the risks posed by foreign bad actors gaining access to advanced AI capabilities. Researchers from Gryphon Scientific and Rand Corporation pointed out that advanced AI models can provide useful information for creating biological weapons.
In its 2024 Homeland Threat Assessment, the Department of Homeland Security says cyber attackers will use AI to “develop new tools” that will “enable larger, faster, more efficient, and more evasive cyber attacks.” He said it is likely.
“The possibility of an explosion is [AI’s] “The use and exploitation is extreme, and we are really having a very hard time complying with it,” Brian Holmes, an official in the Office of the Director of National Intelligence, said at an export control meeting in March, calling China's progress a special I flagged it as a problem. concern.
Crackdown on AI
To address these concerns, the United States has taken steps to prevent American AI chips and the tools to make them from leaving China.
It also proposed rules that would require U.S. cloud companies to notify the government if foreign customers use their services to train powerful AI models that could be used in cyberattacks. .
However, so far there has been no mention of the AI model itself. Alan Estevez, who oversees U.S. export policy at the Department of Commerce, said in December that the department is considering options to regulate open source large-scale language model (LLM) exports before seeking industry feedback. Stated.
Tim Fist, an AI policy expert at the Washington, D.C.-based think tank CNAS, said the threshold is “a good temporary measure until we develop better ways to measure the capabilities and risks of new models.” says.
Jamil Jaffer, a former White House and Justice Department official, said the Biden administration should choose controls based on the model's capabilities and uses, rather than using thresholds for computing power. “Focusing on national security risks rather than technology breaking points is a better strategy because it's more persistent and threat-focused,” he said.
The threshold is not fixed. One of the people said a combination of other factors, such as the type of data and the potential uses of the AI model (such as the ability to design proteins that could be used to make biological weapons), ultimately led Commerce to reach a lower bound. There is a possibility that it will fall below.
Regardless of the threshold, it is difficult to control the export of AI models. Many models are open source, meaning they remain outside the scope of export controls under consideration.
Even imposing regulations on more advanced proprietary models will prove difficult, as regulators will likely struggle to define appropriate criteria for deciding which models to regulate, Fist said. pointed out that China is likely to be only about two years behind the United States. Developed original AI software.
The export restrictions being considered would affect access to the back-end software that powers some consumer applications, such as ChatGPT, but would not restrict access to the downstream applications themselves.
Additional reporting by Karen Freifeld and Anna Tong.
