California Committee Advances Major AI Legislation But with Changes
Key highlights this week:
We’re currently tracking 678 bills in 45 states related to AI this year, 80 of which have been enacted into law.
California lawmakers advanced major AI legislation out of committee last week with some changes to address concerns raised by the industry.
An Indiana task force discussed the use of AI in state government.
New Mexico lawmakers heard from three legislators from other states on their efforts to regulate data privacy and AI.
The AI industry is nervously monitoring California for what could be a new standard in regulation of the nascent technology. California lawmakers advanced major AI legislation this week, but with changes to address concerns raised by developers. The amendments may smooth the path for legislation to pass before the session adjourns later this month.
As originally written, the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act (CA SB 1047) was designed to implement a regulatory regime to certify models compliant with regulations that mitigate the risk of critical harm to consumers. The bill was amended in June to focus the scope on very large models, clarify liability for derivative models, and address concerns by open-source model proponents over requirements for full shutdown capabilities.
Last week, the bill was amended again and passed out of the Assembly Appropriations Committee to the floor, where it could get a full vote. The biggest change is the bill will no longer create a new government agency to oversee the regulation of AI systems. The bill no longer creates the proposed Frontier Model Division, although it still creates the Board of Frontier Models to issue guidance, set thresholds, and regulate audits. The bill instead assigns oversight to the California Government Operations Agency with compliance statements and notice of incidents to be submitted to the Attorney General’s office.
The amended bill narrows the scope of enforcement by the attorney general, limiting civil actions to violations that cause death or bodily harm, harm or theft to property, or that constitute an imminent risk or threat to public safety, or for preventing employees from disclosing information to the attorney general. The attorney general can also seek injunctive relief, but the amendments focus liability on events after they have happened, rather than before, even removing the penalty of perjury for statements submitted by developers for compliance.
Developers are no longer required to provide “reasonable assurance” the model will not create unreasonable risks, and instead are required to “take reasonable care” to avoid producing a model that creates an unreasonable risk or materially enables a critical harm. Last week’s changes also amend the definition of “covered model” to clarify liability for developers. The amendment exempts fine-tuned covered models that cost less than $10 million, a nod to open-source software advocates.
Some AI developers have opposed the bill, arguing instead for a federal solution. The bill’s sponsor, Senator Scott Wiener (D), agrees on the need for a federal solution, but argues there is a need for state regulation while Congress is idle on the issue. Critics still argue the bill would hinder open-source AI software, although Wiener counters the bill has been amended to “accommodate the unique needs of the open-source community.” The computing thresholds have also raised concerns the bill could stifle startups.
The measure is expected to pass the Legislature before the scheduled adjournment date of August 31, giving Governor Gavin Newsom (D) until the end of September to sign or veto the bill. He has not indicated his support yet, but earlier expressed concern about overregulating AI, and lobbyists have focused their efforts on persuading him to use his veto pen. Members of the California Congressional delegation, including Rep. Nancy Pelosi (D), have also voiced opposition.
The bill has not reached the finish line yet, but California could soon join Colorado and Utah as the only states with comprehensive AI regulation. As the legislative session draws to a close, the final outcome will not only shape California's approach to AI regulation but could also set a precedent for how other states, and potentially the federal government, address the complexities of this rapidly evolving field.
Recent Developments
In the News
AI For Mayor: On Tuesday, Cheyenne, Wyoming mayoral candidate Victor Miller lost the election after vowing to exclusively govern the city through an AI chatbot he calls “VIC”, short for “VIrtual Integrated Citizen.” VIC had previously been shut down by OpenAI for violating policies against using its products for campaigning.
San Francisco: The city attorney sued 16 AI-powered “undressing” websites that create deepfake nude images of women without their consent. The lawsuit alleges violations of California and federal law prohibiting the nonconsensual distribution of sexual images.
Major Policy Action
California: On Wednesday, California struck a first-in-the-nation deal for the state to pay tech companies to pay for journalism and AI research. Under the deal, the state and tech companies would pay nearly $250 million to support California-based news organizations and to launch a National AI Accelerator.
Indiana: The Artificial Intelligence Task Force held their first meeting on Wednesday to discuss AI use in state government and data governance. There are no future meetings currently scheduled, but the task force plans to discuss guidance for state use AI and potential costs.
New Mexico: Earlier this month, the Courts, Corrections & Justice Committee heard from state lawmakers from three other states about their efforts regulate AI and data privacy. Connecticut Senator James Maroney, Maryland Senator Sar Lova, and Minnesota Representative Steve Elkins shared their experiences pushing for data privacy protections. The panel also heard from academics on AI discrimination.