A new California law passed earlier this week creates a framework for “Independent Verification Organizations” that can test frontier AI models before they are released, thereby limiting any risks those models pose.
The new law comes as the feasibility of AI oversight is in question for being impossible and too expensive. Laws like this one, even when they are passed with relative haste, become outdated almost immediately.
The models built by frontier labs have become so vast that they require other AI models to analyze them. The METR investigation into the OpenAI-Hugging Face hack required the equivalent of $400,000 in AI tokens — a bill covered by OpenAI — and that was just one incident.
Rather than bills that seek to create frozen-in-time fixes for fast-moving targets, what’s needed is a more nimble approach to technological governance across the board. We need rules and best practices created by smart people who aren’t afraid to be wrong and course-correct. Otherwise, we’ll end up with reams of ineffectual laws or, worse, ones that create the opposite of the intended outcome — something Californians know all about.