After Governor Gavin Newsom vetoed SB 1047 in September 2024, many observers assumed California's appetite for frontier-AI legislation had cooled. Instead, the state regrouped. In late 2025, Newsom signed SB 53 — the Transparency in Frontier Artificial Intelligence Act — authored once again by Senator Scott Wiener. As of 2026, the law is in force, making California the first major US jurisdiction to impose binding transparency obligations on developers of the largest AI models.
SB 53 deliberately avoids the most contested feature of its predecessor: a private right of action and pre-deployment liability for catastrophic harms. In its place sits a comparatively modest disclosure-and-reporting regime. For policymakers across the United States — and for AI labs trying to plan compliance — the contrast matters. SB 53 is a working test of whether transparency, rather than liability, can become the operating model for US frontier-AI governance.
What the Law Actually Requires
The statute applies to a narrow set of developers: those training models above a defined compute threshold. The intent, mirroring the federal Executive Order framework and the EU AI Act's general-purpose-AI tier, is to leave smaller labs, academic researchers, and most start-ups outside the perimeter. Covered developers must:
- Publish a safety and security framework describing how they evaluate and mitigate catastrophic risks before deploying frontier models;
- Report critical safety incidents to California's Office of Emergency Services (CalOES) within statutory timeframes;
- Maintain whistleblower protections for employees and contractors who raise safety concerns internally or with regulators.
Notably, SB 53 does not authorize CalOES or the Attorney General to block model releases, dictate model architectures, or require pre-market approval. The agency's role is principally to receive, aggregate, and where appropriate publish information — not to license technology.
Why This Is the Right Posture, For Now
The case for SB 53's restrained design rests on three propositions that the pro-innovation policy community has long advanced.
First, evidence beats anticipation. Frontier AI capabilities are evolving on a quarterly cadence. Locking in liability rules or capability bans today risks regulating against a snapshot of the technology that will look unrecognizable in eighteen months. Mandatory disclosure builds the empirical record regulators will need before reaching for harder tools. The Stanford AI Index has documented for several years how thinly evidenced many catastrophic-risk claims still are; an enforceable transparency baseline begins to fix that.
Second, transparency is rivalrous with secrecy, not with innovation. Requiring a published safety framework imposes real but bounded compliance costs. It does not prescribe what counts as “safe enough,” nor does it expose proprietary weights or training data. Major labs — Anthropic, OpenAI, Google DeepMind, Meta — already publish responsible-scaling or preparedness frameworks voluntarily. SB 53 turns a competitive norm into a floor without dictating the ceiling.
Third, whistleblower protection is the cheapest accountability mechanism we have. Internal employees see safety problems long before regulators do. The 2024 open letter from current and former frontier-AI staff calling for a “right to warn” underscored how thin existing protections were. Codifying them imposes negligible burden on responsible firms while sharply raising the cost of suppressing bad news at irresponsible ones.
Where Risks Remain
None of this means SB 53 is costless. Three risks deserve close monitoring as implementation proceeds.
The compute threshold will need updating. Training-compute proxies for capability are already a leaky abstraction — algorithmic efficiency gains and inference-time scaling can produce frontier-grade behavior at lower training FLOPs. If the threshold is not revisited regularly, either too many or too few models will fall in scope.
The “critical safety incident” definition will be tested in practice. Regulators should resist scope creep that turns the reporting channel into a generalized AI-harms hotline. The narrower the trigger, the more useful the resulting dataset.
Finally, federal pre-emption looms. A patchwork in which California, New York, Colorado, and Texas each define frontier-AI obligations differently would impose real deadweight costs on developers and ultimately on users. SB 53's relatively light touch — and its explicit framing as a transparency rather than liability regime — should make it easier to harmonize with any future federal framework, whether through NIST, the AI Safety Institute, or Congress.
A Template Worth Borrowing
Other states drafting AI legislation in 2026 should study SB 53 carefully. The instinct to legislate something — anything — on AI is politically powerful and not always wise. California has chosen to require disclosure, protect whistleblowers, and gather data, while leaving substantive safety judgments to the developers closest to the technology. That is a defensible balance: it acknowledges that frontier AI presents genuine systemic questions without pretending the state has the technical capacity to answer them ex ante.
The pro-innovation position is not anti-regulation. It is anti-premature regulation. SB 53 is a credible attempt to thread that needle. Whether it holds up will depend less on the statute itself than on how CalOES interprets its mandate, how the legislature updates the compute threshold, and whether Congress eventually steps in with a coherent federal baseline. For now, California has produced something rare in AI policy: a law that is plausibly proportionate to what we actually know.