A report by the National Telecommunications and Information Administration (NTIA) recommends that the U.S. government should not regulate or restrict the release of open-source AI models at this time.
Instead, it recommends continually assessing the risks and benefits of open-source AI and intervening only when necessary. The report focuses on "dual-use foundation models," defined in a 2023 presidential order as AI models with at least 10 billion parameters trained for a wide range of applications that could pose serious risks to public safety and health, such as facilitating the spread of weapons of mass destruction.
According to the NTIA report, there isn't enough evidence currently to justify restrictions on open-source models. Banning the publication of all or parts of dual-use foundation models now would limit the gathering of important data while preventing researchers, regulators, civil society, and industry from learning more about the technology.
The agency suggests that the government should instead monitor the risks and opportunities of open-source AI models, and build the capacity to respond quickly if needed. The report notes that the cost-benefit analysis could change over time.
Open weight models may be riskier, but they drive innovation and accessibility
In particular, the report examines "open-weight" models that developers can adapt flexibly to create AI tools for small businesses, researchers, nonprofits, and individuals. But they are potentially the most dangerous for the same reason. NTIA's recommendations aim to foster innovation and broaden access to this technology.
To keep track of emerging risks, the report calls for the US government to develop an ongoing program to gather evidence of risks and benefits, evaluate that evidence, and act on it. This could include potential future restrictions on the availability of model weights if warranted.
The report is designed to enable the US government to respond swiftly to risks that may arise from future models. The NTIA acknowledges that restrictions on open-source models could potentially be useful in the future.
