US Prepares to Limit Export of Advanced AI Models to Protect Against Chinese Threat

The Biden administration is planning to impose restrictions on the export of advanced artificial intelligence (AI) models, known as AI Models, which are the software at the heart of AI systems like ChatGPT. This comes as part of a wider effort to protect US AI technology from China, which has made significant strides in this area. The Commerce Department is reportedly considering a new regulatory push to limit the export of proprietary or closed source AI models, whose software and training data are kept confidential. This follows previous measures aimed at preventing the export of high-end AI chips to China, which were put in place over the past two years in an attempt to curb Beijing’s development of these technologies for military purposes. However, some experts warn that it will be challenging to keep up with the fast-moving nature of AI innovation. Current rules allow US AI giants like Microsoft-backed OpenAI, Alphabet’s Google DeepMind, and rival Anthropic to sell their models to clients worldwide without any regulation. This has raised concerns about the potential misuse of these technologies by US adversaries, particularly given their potential applications in fields such as cyber attacks and the development of biological weapons. To address these issues, the administration is reportedly mulling over using a threshold measure based on the computing power required to train an AI model as a means of determining which models should be subject to export restrictions. If implemented, this would affect the export of AI models that have yet to be released, as it is believed that no current models meet the necessary computational requirements. The administration is, however, still at the consultation stage and has not yet finalized any proposals. The move comes amid growing concerns within the US intelligence community, academia, and think tanks about the potential risks posed by foreign powers gaining access to advanced AI capabilities. These include the potential use of AI to facilitate large-scale, fast, efficient, and more evasive cyber attacks, as well as the possibility of AI being utilized to aid in the creation of biological weapons. Any new export rules could also affect other countries. Brian Holmes, an official at the Office of the Director of National Intelligence, expressed particular concerns regarding China’s advancements in this area during an export control meeting held earlier this year. As part of the administration’s broader efforts to counteract China’s AI ambitions, the US has also taken steps to prevent the export of advanced AI chips and associated production equipment to China. The administration has additionally proposed a rule requiring US cloud service providers to inform authorities whenever foreign entities use their platforms to train powerful AI models that could potentially be employed in cyber attacks. However, while the threshold approach is currently being considered, Tim Fist, an AI policy expert at Washington DC-based think tank CNAS, believes that further refinement may be required, particularly as China is believed to be approximately two years behind the US in terms of AI software development. Restrictions on AI model exports are expected to be challenging to implement given the open-source nature of many such models, leaving regulatory bodies struggling to determine which models should be subject to export controls. Furthermore, the rapid pace of technological innovation in this area is likely to present ongoing challenges for regulators attempting to maintain effective controls.