The state of California has registered a draft law called the Safe and Secure Innovation for Advanced Artificial Intelligence Models Act (SB-1047). This project focuses on the existential fears of future AI models and may significantly limit research and development in this area. This was reported by Ars Technica.
The bill was introduced by state Senator Scott Wiener, passed by the California Senate in May by 32 to 1, and seems likely to be put to a final vote in the state Assembly in August 2024.
The text of the draft law requires companies developing artificial intelligence models to introduce procedures and testing systems to prevent and respond to "security incidents."
The project provides a legal definition of these security incidents that focuses on defining a set of "critical damages" that an AI system can cause. This includes damage that results in "mass casualties or losses of at least $500 million."
The law provides several examples of such incidents: "the creation or use of chemical, biological, radiological or nuclear weapons" and "precise instructions for a cyberattack on critical infrastructure."
To prevent these, frankly, sci-fi situations, anyone training a large enough model should "implement a quick shutdown capability" and have a policy in place in case of such a shutdown, among other precautions and tests.
The bill is authored by Dan Hendricks, co-founder of the Center for AI Security (CAIS). In an article published in Time Magazine, Hendricks notes that "evolutionary pressures are likely to instill in AI self-preserving behaviors" and lead to "the displacement of humanity as the dominant species on Earth."
The full text of the project can be found here.