The U.S. Department of Commerce today released a report supporting “open-weighted” generative AI models like Meta's Llama 3.1, but recommending that the government develop “new capabilities” to monitor these models for potential risks.
The report, written by the Commerce Department's National Telecommunications and Information Administration (NTIA), argues that open-weighted models would broaden the scope of generative AI available to small businesses, researchers, nonprofits, and individual developers. For these reasons, the report suggests, governments should not restrict access to open models, at least not before studying whether such restrictions would have a negative impact on the market.
This sentiment echoes recent comments from FTC Chairwoman Lina Khan, who believes an open model will allow more small businesses to bring their ideas to market, thereby promoting healthy competition.
“The openness of the largest and most powerful AI systems will impact competition, innovation, and risk in these transformative tools,” Alan Davidson, Under Secretary of Commerce for Communications and Information and NTIA Administrator, added in a statement. “NTIA's report recognizes the importance of open AI systems and calls for more aggressive oversight of the risks posed by having model weights from the largest AI models widely available. The government has a critical role to play in supporting AI development while building the capacity to understand and address emerging risks.”
The report comes as domestic and international regulators are considering rules that could restrict or impose new requirements on companies wanting to launch open-weight models.
California is on the verge of passing SB 1047, which would require all companies that train models using more than 1026 FLOPs of computing power to strengthen cybersecurity and develop ways to “shut down” copies of the models under their control. Abroad, the EU recently finalized compliance deadlines for companies under its AI law, which imposes new rules on copyright, transparency, and AI applications.
Meta has said that EU AI policies will prevent it from releasing some of its open models in the future, and many startups and large tech companies have spoken out against California's law, arguing it is too strict.
NTIA’s model governance philosophy is not entirely laissez-faire.
In its report, NTIA calls on the government to develop an ongoing program to collect evidence of the risks and benefits of open models, evaluate that evidence, and act on that evaluation, including by imposing certain limitations on the availability of the models, if necessary. Specifically, it suggests that the government examine the safety of different AI models, support research into risk mitigation, and develop “risk-specific” indicator thresholds that indicate policy changes may be necessary.
Commerce Secretary Gina Raimondo noted that these and other measures are in line with President Joe Biden's executive order on AI, which calls for government agencies and companies to set new standards for the development, deployment and use of AI.
“The Biden-Harris Administration is taking every step to maximize AI's potential while minimizing its risks,” Raimondo said in a press release. “Today's report provides a roadmap for responsible AI innovation and American leadership by embracing openness and recommending ways the U.S. government can prepare for and adapt to potential challenges ahead.”