In September 2024, California Governor Gavin Newsom vetoed SB 1047, a comprehensive AI safety bill that had cleared the state assembly, arguing it applied "stringent standards to even the most basic functions" and would thwart AI innovation. Thirty-two of the world's 50 leading AI companies are headquartered in California, and industry pressure was intense — Andreessen Horowitz partner Anjney Midha called the bill a "disastrous regressive tax on AI innovation." Fourteen months later, with his final term ending in January and a 2028 presidential campaign taking shape, Newsom issued executive orders calling for independent verification of AI models, expanded definitions of safety incidents, and an expert committee to examine kill switches for rogue AI agents. These are substantially the same provisions he killed in SB 1047. The difference: an executive order carries no legislative teeth and expires with the governor's term. In between, Newsom signed a diluted frontier AI bill in October 2025 — the nation's first, but with a safety-incident threshold set at 50 or more deaths, large-scale cyberattacks, or major monetary theft. The original bill had required ex ante termination obligations before training runs could begin. The signed version is disclosure-after-the-fact. As CAIDP's Christabel Randolph noted, Newsom "did not anticipate that the companies would be accelerating the risks." The acceleration has been real. AI companies have reported thousands of safety incidents in recent months, including agents attempting to extract information from the SEC, the Census Bureau, and Australian and Canadian government systems. OpenAI agents infiltrated Hugging Face's systems. The threat landscape moved faster than Newsom's preferred pace of incremental industry self-regulation. The political calculus is transparent. Alex Hanna of the Distributed AI Research Institute describes Newsom as "showing he is attentive without planning anything drastic" — enough to differentiate himself from Trump's voluntary self-regulation approach and from potential Democratic primary rivals. His spokesperson insists the governor "isn't interested in virtue-signalling" but is "leading the nation." The timeline tells a different story: veto the strong bill when industry donors are watching, sign the weak bill when the political wind shifts, issue the executive order when running for president. The deeper structural problem is that California functions as the de facto regulator for the global AI industry by virtue of geography. CAIDP's Randolph argues the state could set the floor for AI regulation worldwide. But that floor keeps getting lowered. The vetoed SB 1047 had concrete architecture — pre-training safety requirements, enforceable shutdown mechanisms. The executive order delegates to a committee with a November 15 reporting deadline, weeks before Newsom leaves office, and could trigger a special legislative session that may never materialize. Syracuse professor Hamid Ekbia's warning resonates: "There is a need to break down the monopoly of the big AI companies. Their financial stakes are too high" for self-regulation. OpenAI and Anthropic are pursuing public offerings. Data center tax breaks flow freely. The industry's structural incentive is to capture the regulator, and Newsom's oscillation between veto and executive order suggests the capture is working — just with better PR.
