On Wednesday, December 13, 2023, the Department of Commerce’s National Telecommunications and Information Administration (NTIA) launched a public engagement effort related to its review of the risks and benefits of certain highly-capable artificial intelligence (AI) models. Among the issues that NTIA will examine to assess these risks and benefits are the model weights that “open” (i.e., publicly available) AI models use to generate model outputs.
As the NTIA explained it, “[m]odel weights govern how AI algorithms evaluate data by providing instructions as to which factors matter a great deal and which matter less or not at all. Machine language algorithms assess large volumes of data and learn – or are taught – which relative weightings of the many data points available will lead to the most accurate determinations. If the model weights are widely available, the model itself can become easily replicable or potentially manipulable.”
President Biden’s recent Executive Order on AI safety and security requires the Department of Commerce to solicit public input on policy and regulatory approaches to such open models and submit a report to the President based on that input. In keeping with the Executive Order, NTIA’s review will focus on “dual-use foundation models,” powerful models which, as NTIA explained in its press release, “are generally useful for many purposes.” NTIA will consider the risks associated with actors’ modification of open models, including by removing safeguards; the benefits of open models to research and innovation; and potential regulatory mechanisms that could manage risks while maximizing benefits.
NTIA has stated that it intends to issue a formal Request for Comment on these issues in early 2024.