You know what’s a good idea for “pacing” the frontier of AI development? Laws! There are an increasing number of them not only being drafted but also being passed by state governments. California was first with a law focused specifically on frontier AI model risk, which is already in effect. New York followed suit, and most recently Illinois passed its AI Safety Measures Act. These laws are precisely targeted at the kinds of catastrophic risks that AI developers are wringing their hands over.
California’s AI whistleblower protections may even already be yielding benefits as they offer protection to the likes of Jacob Coxon, who recently resigned from Anthropic over AI fears. Perhaps others will also be emboldened to come forward. The laws from California, New York, and Illinois create obligations for publishing frontier AI safety frameworks (tracked here), as well as other bits of transparency. Just the fact that Anthropic and OpenAI have made explicit disclosures of their safety frameworks in compliance with California’s law suggests that their models were trained with more than 10^26 compute operations — otherwise they wouldn’t need to make that disclosure. Some are already raising questions about OpenAI’s compliance with the California law — do recent cybersecurity incidents signal a real loss of control?
The California, New York, and Illinois laws are largely similar. But they also seem to be building on each other. Illinois, for instance, includes provisions for independent third-party audits of compliance with the law. That’s a powerful addition that enhances accountability and was something that was in an early version of the New York law, but got stripped from the final. An advance in both the New York and Illinois laws is the addition of registration and ownership disclosure requirements. For the sake of human responsibility and accountability, it’s critical that society know who the beneficial owners of these large companies are.
These laws aren’t perfect. They are narrowly scoped and don’t address many of the current harms already being enacted by AI systems in society. The degree of similarity in them suggests they are probably being influenced by the tech lobby. They also aren’t responsive enough to deal with the rapid pace of AI advances, and there’s a lag with when they are signed into law and come into force which creates a gap just when we need information the most. But they’re something to build on, and they give us some transparency in the form of safety frameworks (i.e. prevention plans) that the frontier AI companies can and should be held to until the federal government decides to act seriously to regulate AI.