Which organisations does SB 942 actually cover?

Covered providers of generative AI systems producing image, video or audio content, with more than one million monthly users, publicly accessible in California. Text-only systems and systems below the user threshold fall outside the covered-provider obligations. AB 853 separately phases in duties for large online platforms from January 1, 2027 and for capture-device manufacturers from January 1, 2028, so an organisation outside scope today may be inside scope later.

When does the 96-hour licence revocation clock start?

At the covered provider's discovery that the licensee no longer maintains the capability to include a latent disclosure. Discovery is an event the provider must be able to date, which makes the monitoring schedule and the discovery record compliance evidence in their own right. A provider that revoked promptly but cannot show when discovery occurred has a weaker position than the facts deserve.

Does the August 2, 2026 date replace January 1, 2026 everywhere?

For the covered-provider provisions, yes. AB 853 moved the operative date to August 2, 2026. The later phase-in dates for large online platforms and capture-device manufacturers are separate and unchanged by that shift. Internal trackers and control descriptions carrying the original January 1, 2026 date should be corrected, because a stale date in a compliance artefact reads as an unmaintained programme.

Where does DeepInspect fit against SB 942?

DeepInspect evaluates HTTP traffic between authenticated users or agents and LLM endpoints routed through it, and records identity, model destination, policy version and outcome per decision. Latent disclosure embedding, manifest disclosure rendering, and the public detection tool are model-layer and application-layer functions outside that boundary. The relevant contribution is the deployer-side record of which identity requested generation under which policy, and when.

California SB 942 AI Incident Reporting: The 96-Hour Licensee Clock

The California AI Transparency Act contains one deadline that behaves more like an incident response requirement than a disclosure rule. A covered provider that discovers a licensee no longer maintains the capability to include a latent disclosure has 96 hours to revoke the licence. Four days, running from discovery.

A duty that starts at discovery is a duty that depends on detection, and detection is where most programmes around this statute are thinnest.

TL;DR

SB 942, enacted September 19, 2024, applies to covered providers of generative image, video and audio systems with more than one million monthly users.
AB 853, signed in October 2025, moved the operative date from January 1, 2026 to August 2, 2026, and phased in obligations for large online platforms from January 1, 2027 and capture-device manufacturers from January 1, 2028.
Covered providers must offer latent and manifest disclosures and a publicly available AI detection tool.
A licensee that stops maintaining latent disclosure capability must have its licence revoked within 96 hours of the provider discovering it.
The civil penalty is $5,000 per violation, which makes the per-item counting basis the number worth modelling.

What the statute requires, with the amended dates

SB 942 applies to covered providers of generative AI systems producing image, video or audio content with over one million monthly users, publicly accessible in California. The obligations are a latent disclosure embedded in the content, a manifest disclosure option available to users, and a publicly accessible detection tool that lets anyone check whether content came from the provider's system.

AB 853 changed the timetable. The covered-provider provisions became operative on August 2, 2026. The original date was January 1, 2026. Large online platform obligations phase in from January 1, 2027, and capture-device manufacturer obligations from January 1, 2028.

Any internal document still carrying the January 1, 2026 date is stale, and it is the most common error I see in state AI law trackers maintained in spreadsheets.

The 96-hour duty is a detection problem

The licensee provision is the operative incident requirement. Licensees must maintain the capability to include a latent disclosure in content the licensed system generates. Where a covered provider discovers that a licensee no longer maintains that capability, the provider must revoke the licence within 96 hours of discovery.

Everything in that sentence hinges on the word discovers. A provider with no monitoring discovers non-compliance when a journalist calls. A provider with sampling in place discovers it during a scheduled check and has a dated record showing when the clock started.

Three artefacts make the duty operable. A schedule of licensee capability checks, with dates and method. A discovery record naming what was found, by whom, and at what timestamp. A revocation record showing the action and the elapsed time. Without the middle one, a provider cannot demonstrate compliance with the deadline even when it revoked quickly, because there is nothing establishing when the 96 hours began.

Penalties count per violation

The civil penalty is $5,000 per violation. For a content-marking obligation applied to generated outputs, the counting basis matters more than the headline figure, and a programme modelling exposure per incident instead of per item is likely understating it.

That arithmetic is the argument for making the control preventive rather than detective. A disclosure that failed to attach on ten thousand outputs is a different conversation from one that failed on a handful before a check caught it, and the difference is entirely a function of how quickly the failure was detected.

Our California SB 942 controls mapping piece places these obligations against a control set, and the audit evidence piece covers what to retain.

Where this sits against the enforcement boundary

Most of SB 942 is out of reach for a policy gateway, and that is worth stating before anyone builds a slide claiming otherwise. Embedding a latent disclosure happens inside the generation process at the model or application layer. Manifest disclosure is rendered by the application. The public detection tool is a product a covered provider builds and hosts. None of those three is HTTP traffic between an authenticated user or agent and an LLM.

The slice that is in scope is narrow and real. Where an organisation routes generation requests through a policy boundary, that boundary can record which identity requested generation, against which model, under which policy version, and whether a disclosure requirement was in force at the time. For the deployer side of the regime, that record is the discovery and evidence layer rather than the marking mechanism.

Keeping the distinction explicit is also what makes the compliance argument credible. The EU AI Act comparison work shows how quickly transparency regimes get conflated, and California's statute is narrower than most summaries of it suggest.

Preparing the discovery process

Write the licensee check as a scheduled control with a named owner, a method, and a sampling rate. Generate content through each licensed deployment, verify the latent disclosure is present, and retain the result with a timestamp whether it passed or failed.

Write the escalation path so that a failed check produces a dated discovery record immediately rather than after a triage meeting. The 96 hours runs from discovery, and an internal process that delays the record does not delay the clock.

Then rehearse one revocation. Pick a test licence, run the process end to end, and measure the elapsed time. Four days sounds generous until the revocation requires a contract review, a commercial conversation and an engineering change, at which point it stops sounding generous at all. The AI governance audit framework covers how to hold these records alongside other regimes.

DeepInspect

DeepInspect provides an independent policy decision at the HTTP AI request boundary. For generation traffic routed through it, DeepInspect evaluates the application-supplied identity and role against organisational policy for content classification and model authorisation before the request reaches the model.

Each decision produces a signed, tamper-evident record held outside the calling application's write path. Where a transparency regime turns on when something was discovered and what policy was in force at the time, that record supplies the dated evidence a discovery log is supposed to contain.

Book a technical deep dive at deepinspect.ai.