California’s Advanced AI Act: Stanford and Berkeley recommend broadening the definition of the models covered

Rob Reich and Deirdre Mulligan advise California to move beyond the 10²⁶ FLOP threshold and to introduce safeguards for reports made in good faith.

Rob Reich (Stanford HAI) and Deirdre K. Mulligan (UC Berkeley) have submitted a report to the California Department of Technology (CDT) to inform its first annual review under the Transparency in Frontier AI Act (TFAIA), the state's law on the transparency of advanced AI models. The document summarizes two days of workshops organized on August 12 and 13 with various stakeholders, and formulates recommendations that are binding only on the two authors.

The main proposals

The authors invite the State to no longer define a "cutting-edge" model solely by a calculation threshold greater than 10^26 operations, to broaden the scope of catastrophic risks taken into account, to treat the theft of a model's weights separately from the thresholds for physical damage, to restrict the exception provided for assessments, and to create legal protection for reports made in good faith.