Senate Bill 1047 would require developers to conduct safety and security testing, risk assessments of AI models above a certain size before training or releasing them, as well as annual third-party audits. A committee of the California state assembly passed the legislation Thursday. (The Logic)

Talking point: Bill sponsor Scott Wiener said his proposal is “as light touch and pro-innovation as possible.” Decorated Canadian AI researchers Geoffrey Hinton and Yoshua Bengio backed the bill. But prominent U.S. scientists including Yann LeCun and Fei-Fei Li panned it. Firms like Anthropic and Meta opposed the law’s original text, which they said would make them liable for potential problems rather than actual negative outcomes, and for harms that are the fault of users rather than developers. Investors claimed it would deter new startups. Weiner amended the bill to limit pre-deployment enforcement and cut a proposed new regulator. The big AI  firms have made similar complaints about Canada’s proposed AI law.