California AI Law in 2026: Which Statutes Bind an Enterprise Deployer, and When

Five separate California statutes now touch how an enterprise uses AI, and they carry four different effective dates and three different definitions of who is on the hook. The California AI law a compliance owner actually has to track in 2026 is not one bill. It is SB 1001 (operative since July 1, 2019), AB 2013 (operative January 1, 2026), SB 942 as amended by AB 853 (operative August 2, 2026), the California Privacy Protection Agency's automated-decisionmaking rules under the CCPA (phasing in from January 1, 2027), and SB 53 (operative January 1, 2026). Some of these bind the company that builds a model. Some bind the company that deploys one. The mistake I watch compliance teams make is reading a headline about "California's new AI law," assuming it lands on them, and building a control that answers the wrong statute.

I want to walk each one, name the effective date, and say plainly who it binds. Then I will be honest about the narrow slice of this a policy gateway can help you evidence, and the larger slice it cannot.

SB 1001: bot disclosure, and it has been law since 2019

The oldest of the group is the B.O.T. Act, SB 1001, operative since July 1, 2019. It makes it unlawful to use a bot to communicate with a person in California to incentivize a commercial transaction or influence a vote, without disclosing that the account is a bot. The disclosure has to be clear and conspicuous. The Attorney General enforces it under the Unfair Competition Law, with fines reaching $2,500 per violation.

This one binds whoever operates the customer-facing bot. If your company runs an AI chat agent that sells or upsells to Californians, the disclosure obligation is yours, and it has been for years. A gateway on the model path has nothing to say about it. The disclosure is a UI-layer control, rendered in the interface the consumer sees.

AB 2013: training-data transparency, effective January 1, 2026

AB 2013, the Generative AI Training Data Transparency Act, took effect January 1, 2026. It requires developers of generative AI systems made available to Californians to post documentation of the datasets used to train those systems. The obligation lands on the developer, meaning the party that trained or substantially modified the model, not the enterprise that calls an API.

I have kept this section short on purpose, because we cover the documentation requirements, the "substantial modification" trigger, and the fine-tuning edge cases in a dedicated piece. If you fine-tune an open-weights model in-house, read the AB 2013 breakdown before you assume the developer obligation stops at OpenAI or Anthropic. Fine-tuning can pull you into the developer definition.

SB 942 and AB 853: transparency and watermarking, now August 2, 2026

The California AI Transparency Act, SB 942, was originally set to become operative January 1, 2026. That original date slipped when Governor Newsom signed AB 853 on October 13, 2025, which moved the operative date to August 2, 2026 and expanded the scope. The National Law Review's deadline tracker is a clean reference for the amended timeline.

SB 942 requires covered providers of generative AI systems with more than one million monthly California users to offer a free AI-detection tool and to attach manifest and latent disclosures (visible labels plus provenance metadata) to AI-generated content. AB 853 pulled generative AI hosting platforms, large online platforms, and capture-device manufacturers into the regime. The stated reason for the delay was to line California's provenance requirements up with the EU AI Act Article 50 timeline, which we track in the Article 50 transparency guidance.

Watermarking and content provenance are model-layer and application-layer functions. A stateless proxy sitting on the HTTP request path does not watermark generated images or embed C2PA provenance in outputs. If SB 942 binds you, that engineering work lives in your generation pipeline, not in a gateway.

The CPPA automated-decisionmaking rules: the one most likely to bind a deployer

The regulation a compliance owner should read first is the set the CPPA finalized under the CCPA. On September 23, 2025, the California Office of Administrative Law approved the automated decisionmaking, risk assessment, and cybersecurity audit rules. These rules went into effect January 1, 2026, but the automated-decisionmaking-technology (ADMT) obligations phase in later. Businesses using ADMT to make a "significant decision" about a consumer must comply beginning January 1, 2027, and the pre-use notice, opt-out, and risk-assessment obligations attach from April 1, 2027.

ADMT is defined narrowly, as technology that processes personal information and uses computation to replace or substantially replace human decision-making about a consumer. A "significant decision" covers things like lending, housing, employment, and healthcare access. This regulation binds the business that uses the AI to make the decision. That is the deployer, which is probably you. It is the closest California analog to GDPR Article 22's automated-decision provisions, and it is the one where you will need to produce evidence about which system processed which consumer's data for which decision.

SB 53: frontier-model safety, and it probably does not bind you

SB 53, the Transparency in Frontier Artificial Intelligence Act, was signed September 29, 2025 and became operative January 1, 2026, with additional provisions in 2027. It requires frontier developers to publish a safety framework and transparency reports, report critical safety incidents to the Office of Emergency Services, and it carries civil penalties up to $1 million per violation.

SB 53 binds "frontier developers," a category gated by very large training-compute thresholds. Unless your company is training models at the scale of the largest labs, this one governs your vendors rather than you. Read it to understand what your model provider is now obligated to publish, not because you have a filing to make. My opinion, and a compliance committee would want to soften it: most enterprises spend too much attention on SB 53 because it is the headline-grabbing "AI safety" law, and too little on the CPPA ADMT rules that will actually generate their audit requests.

Sorting binder from bystander

The practical filter for the whole California AI law picture is a single question: does the statute regulate building the model or using the model. SB 942, AB 2013, and SB 53 mostly regulate the building and the publishing of AI systems and content. SB 1001 and the CPPA ADMT rules regulate the using, and those two are where a typical enterprise deployer carries direct obligations. This same builder-versus-user split runs through the Colorado AI Act, Utah's AI disclosure law, and Illinois' AI employment rules. For a snapshot of how these state regimes stacked up this quarter, the Q2 2026 compliance mini-report has the comparison table.

The obligation that produces the most evidence work, across almost all of these, is proving which AI interactions happened, under whose identity, and against which policy at the moment of the decision.

DeepInspect

Here is the narrow slice of California AI law that a policy gateway helps you evidence, stated without inflation. When the CPPA ADMT rules require you to account for the automated processing behind a significant decision, the question a regulator asks is factual: which model was called, by which authenticated user or agent, with what category of personal information in the request, and what policy governed that call. DeepInspect sits on the HTTP path between your authenticated users and agents and the LLM, enforces identity-aware policy on each request and response, and writes a per-decision audit record. That record is contemporaneous evidence of which AI interactions occurred under which identity, which is exactly the raw material an ADMT risk assessment or a regulator's traceability request needs. The audit-trail-by-regulation reference maps this to the specific record fields each regime expects.

I will be equally clear about what the gateway leaves untouched, because overclaiming here is how vendors lose credibility with a CISO. It does not watermark generated content or provide the SB 942 detection tool. Publishing the training-data documentation AB 2013 requires falls to your model developers. Rendering the SB 1001 bot disclosure happens in your product UI, and filing an SB 53 transparency report is a frontier developer's job, not yours. Those are application-layer and model-layer obligations that a gateway on the request path has no way to satisfy. What it gives you is the identity-bound, per-decision record that turns "we think our AI use is compliant" into "here is the log of every AI decision, tied to an identity, that a regulator can inspect." If you are mapping your California obligations to actual controls, book a technical deep dive at deepinspect.ai.