On 2 August 2026, the California AI Transparency Act — enacted as SB 942 and amended by AB 853 — became operative. It did so on precisely the same day the European Union’s AI Act Article 50 transparency obligations began to apply, which is not a coincidence so much as a convergence: two of the world’s largest regulatory blocs decided independently that the first thing they wanted from generative AI was not safety testing or risk classification but a simple, machine-checkable answer to the question did a machine make this?
For compliance teams, the practical consequence is that the provenance problem has moved from the roadmap to the statute book. This is no longer a question of whether to adopt C2PA or watermarking as a matter of trust and safety policy. In California it is a legal obligation carrying $5,000 per violation, with each day of continuing violation counted as a discrete violation.
This article works through who is covered, what the four operative obligations actually require, how the 96-hour licensee revocation duty changes commercial contracting, how the California regime interacts with the EU’s Article 50 duties, and what a defensible compliance posture looks like before the next phase lands on 1 January 2027.
The one-year delay, and why it matters
SB 942 was signed in September 2024 with an original operative date of 1 January 2026. In 2025 the California Legislature amended the Act — principally through AB 853 — to delay operation to 2 August 2026 and to layer in a phased roadmap of additional duties running through 2028.
That delay is worth dwelling on, because it tells you something about how the obligations were received. The original text placed the entire burden on the developers of generative AI systems. The amended structure recognised that provenance is a chain problem: a latent watermark embedded by a model provider is worthless if the platform that distributes the content strips it, and worthless again if the camera that captured the authentic footage never signed it in the first place. So the Legislature extended the regime outward in phases:
- 2 August 2026 — obligations bite on covered providers (the generative AI system developers).
- 1 January 2027 — obligations extend to generative AI hosting platforms and large online platforms.
- 1 January 2028 — obligations extend to capture device manufacturers (cameras, phones, recorders).
Organisations that read the delay as a reprieve and stopped work have lost a year. Organisations that used it to build provenance pipelines are now in a position to satisfy the 2026 duties and to absorb the 2027 platform duties without re-architecting.
Who is a “covered provider”?
The Act defines a covered provider as a person that creates, codes, or otherwise produces a generative AI system that:
- has more than one million monthly visitors or users, and
- is publicly accessible within California.
Three points follow, each of which is routinely misread.
The threshold is on the system, not the company. A large enterprise with a small generative AI product may fall outside the threshold. A small company with a viral image generator may fall inside it. Counting is done at the level of the generative AI system that is publicly accessible.
“Publicly accessible within California” is not the same as “targeted at California.” There is no meaningful geographic carve-out available to a globally available web product. If Californians can reach it without a bespoke commercial arrangement, it is publicly accessible within California. Enterprise-only systems sold under negotiated contracts sit in a genuinely different position; consumer-facing systems do not.
The obligations apply to image, video and audio — not text. This is the single most important scoping fact in the statute and the one most often lost in summary coverage. A text-only large language model interface, however large, is not carrying SB 942 disclosure duties for its text output. The moment that same product generates an image, a video, or synthesised speech, it is.
The four operative obligations
1. A free, public AI detection tool
A covered provider must make available a publicly accessible detection tool, free of charge, that allows a user to assess whether a given piece of content was created or altered by the provider’s generative AI system. The tool must:
- accept user-uploaded content or a URL,
- output any system provenance data detected in that content,
- not output any personal provenance data about the individual who created the content,
- support an API or otherwise be reasonably accessible, and
- collect no more user data than is necessary to operate.
That last constraint is a privacy overlay that compliance teams should note: the detection tool is itself a personal-information processing activity subject to the CCPA. Building a detection endpoint that logs uploads indefinitely to improve the classifier would satisfy SB 942 and violate the CCPA’s data minimisation duty at the same time. Design it to process and discard.
2. Latent disclosure
The provider must include, in AI-generated image, video and audio content, a latent disclosure that is detectable by the provider’s own tool and, to the extent technically feasible, permanent or difficult to remove. The latent disclosure must convey:
- the name of the provider,
- the name and version number of the generative AI system that produced the content,
- the time and date of creation, and
- a unique identifier for the content.
In practice this points squarely at C2PA Content Credentials and equivalent cryptographically signed manifest standards, combined with a robust invisible watermark for the cases where the manifest is stripped. Neither alone is sufficient: manifests survive editing tools that respect them and are trivially removed by tools that do not; watermarks survive re-encoding but carry far less metadata. The statute’s “to the extent technically feasible” language is doing a great deal of work here, and the defensible position is a layered implementation with documented reasoning about where feasibility ends.
3. Optional manifest disclosure
The provider must offer the user the option to include a manifest disclosure — a clear, conspicuous, appropriately sized visible notice that the content is AI-generated, permanent or extremely difficult to remove. Note that this is a user option, not a mandate on the provider to visibly stamp every output. The compliance duty is to offer it and to make it work, not to force it.
4. Licensee flow-down and the 96-hour revocation clock
This is the provision that changes commercial practice. Where a covered provider licenses its generative AI system to a third party, it must contractually require the licensee to maintain the system’s disclosure capability. And if the provider knows that a licensee has modified the system such that it no longer includes the required disclosures, the provider must revoke the licence within 96 hours.
Ninety-six hours is four days. It is a shorter clock than most breach notification regimes and it runs on knowledge, not on confirmation, investigation, or the conclusion of a commercial negotiation. Compliance teams should treat this as an incident response obligation and build it accordingly:
- a defined intake channel for reports that a licensee has stripped disclosures,
- a documented knowledge trigger and timestamp,
- a pre-approved revocation pathway that does not require a fresh legal review each time,
- and a contractual right of immediate suspension in every model licence.
Renegotiating a licence agreement to insert a suspension right takes longer than 96 hours. The contractual work has to be done in advance.
How this interacts with the EU AI Act’s Article 50
Article 50 of the EU AI Act, applicable from the same date, imposes an overlapping but not identical set of transparency duties:
| California SB 942 | EU AI Act Article 50 | |
|---|---|---|
| Modalities | Image, video, audio | Text, image, video, audio |
| Provider marking | Latent disclosure required | Machine-readable marking of synthetic content required |
| Visible labelling | User option (manifest) | Deployer duty to disclose deepfakes and AI-generated public-interest text |
| Chatbot disclosure | Not covered | Required — users must be told they are interacting with AI |
| Detection tool | Required, free, public | Not required |
| Threshold | >1M monthly users | No user threshold |
| Penalty | $5,000 per violation per day | Up to €15M or 3% of global turnover |
The practical takeaway is that neither regime subsumes the other. California requires a detection tool that Europe does not. Europe reaches text and chatbot interactions that California does not. Europe has no user-count threshold, so a small European-facing product carries Article 50 duties while escaping SB 942 entirely.
Organisations serving both markets should build to the union of the two: machine-readable provenance across all synthetic modalities including text, visible deepfake labelling, chatbot disclosure at the start of every session, and a free public detection endpoint. That superset satisfies both and avoids maintaining two divergent pipelines. Our earlier analysis of the EU AI Act’s 2 August Article 50 obligations covers the European half of that union in detail.
Enforcement: who brings the action
Enforcement authority sits with the Attorney General, city attorneys, and county counsel. The penalty is $5,000 per violation, and — critically — each day a violation continues is a separate violation.
Compliance teams should model what that arithmetic actually means. A systemic failure to embed latent disclosure is not one violation; a plausible reading is that it is one violation per non-compliant output, per day. Even on the most conservative reading — one violation per day for the system as a whole — a six-month gap between the operative date and remediation is $900,000. On a per-output reading the numbers become unbounded very quickly.
There is no cure period written into the statute. There is also, as of this writing, no enforcement action on the public record. That is normal for a regime three weeks old and should not be mistaken for tolerance; California’s privacy regulators have shown across the CCPA enforcement wave and the Delete Act that a quiet first quarter is followed by a sustained cadence of actions.
A compliance checklist
Scoping
- Inventory every generative AI system your organisation produces that is publicly accessible in California.
- Determine monthly visitors/users per system, and document the counting methodology. The threshold determination is the first thing an enforcement inquiry will test.
- Classify output modalities per system. Image, video and audio are in scope; text alone is not.
- Separately scope for the EU AI Act, where there is no threshold and text is in scope.
Latent disclosure
- Implement a signed provenance manifest (C2PA Content Credentials or equivalent) carrying provider name, system name and version, creation timestamp, and unique content identifier.
- Layer an invisible watermark for resilience against manifest stripping.
- Document technical feasibility limits and the reasoning behind them. “To the extent technically feasible” is a defence only if you can show the analysis.
- Verify the four required data elements are actually present. A manifest that omits the system version number is non-compliant even if everything else is right.
Detection tool
- Publish a free, public detection endpoint accepting uploads and URLs.
- Return system provenance data; suppress personal provenance data.
- Provide API access.
- Apply CCPA data minimisation to the endpoint: process and discard, do not retain uploads for model improvement without a lawful basis and disclosure.
Manifest disclosure
- Offer users the option of a visible AI-generated label.
- Ensure the label is clear, conspicuous, appropriately sized, and difficult to remove.
Licensing
- Amend all generative AI licence agreements to require maintenance of disclosure capability.
- Insert an immediate suspension and revocation right exercisable without notice period.
- Stand up a 96-hour revocation runbook with a defined knowledge trigger, named owner, and evidence log.
- Build monitoring for licensee output that would evidence stripped disclosures.
Governance and evidence
- Assign a named accountable owner for SB 942 compliance.
- Log provenance embedding rates as a monitored control with alerting on drops.
- Retain evidence of detection tool availability and uptime.
- Diary the 1 January 2027 hosting-platform and large-online-platform obligations, and the 1 January 2028 capture device obligations.
The underlying shift
It is tempting to file SB 942 alongside the growing pile of state AI statutes and treat it as one more disclosure obligation. That undersells it. What California and the EU have jointly done, on the same day, is convert content provenance from a voluntary industry initiative into infrastructure that carries statutory penalties for absence.
The organisations that will struggle are not the ones that disagree with the policy. They are the ones that treated provenance as a trust-and-safety feature owned by a product team, with no compliance owner, no control testing, and no evidence trail. A watermarking pipeline that works 94% of the time is a good product feature and a $5,000-per-day compliance failure.
The 96-hour licence revocation duty is the clearest signal of legislative intent here. California has decided that provenance is a chain obligation and that the party at the top of the chain is responsible for policing the parties below it — on a clock measured in hours, not quarters. That is the same architecture that GDPR Article 28 applies to processors and that NIS2 applies to supply chains. Generative AI has now joined the list of things regulators expect you to be able to prove, in writing, about your vendors and your licensees.
This article is provided for informational purposes only and does not constitute legal advice. Organisations should consult qualified counsel regarding their specific compliance obligations under the California AI Transparency Act, the EU AI Act, and applicable state and international law.
