California AI Transparency Act: provenance rules for platforms from 1 January 2027

Summary. From 1 January 2027, the California AI Transparency Act (Business and Professions Code sections 22757 to 22757.6) reaches large online platforms: public-facing social media, file-sharing and mass messaging platforms and stand-alone search engines that exceeded 2,000,000 unique monthly users in the preceding 12 months. They must detect provenance data associated with content they distribute, show users whether system provenance data or a digital signature identifies content as AI-generated, AI-altered or captured by a camera or recorder, let users inspect that data, and not knowingly strip it on upload, distribution or download. The same day, GenAI hosting platforms may not knowingly offer models that skip the act's latent disclosures. Two laws signed on 30 September 2026 changed the act. AB 2713 (Chapter 856) rewrote the platform section, with its new text applying from 1 January 2027. SB 1000 (Chapter 861), an urgency statute in force since that day, removed the 1,000,000-user threshold for GenAI providers, dropped the manifest disclosure option and narrowed the entertainment exemption to videogames. Penalties remain $5,000 per violation, with each day a separate violation. Capture-device makers follow on 1 January 2028. Platform teams have 50 days to ship detection, display and preservation.

About this article. For product, platform and media-pipeline engineers and the counsel who advise them. We checked every date, definition and duty against the chaptered bill texts and code text published by the California Legislature, the Governor's signing announcements and the C2PA specification, on 12 November 2026. This is general information, not legal advice: your obligations depend on your facts, so confirm them with California counsel. No vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What changed, and when each duty starts

SB 942 created the act in 2024 as Chapter 291 of the 2024 statutes. AB 853, signed on 13 October 2025 as Chapter 674, moved its start from 1 January 2026 to 2 August 2026 and added the platform, hosting and capture-device duties. In 2026 the Legislature amended it twice more, and the Governor signed both bills on 30 September 2026.

Two of the 2026 changes alter how you scope compliance.

First, SB 1000 deleted the user threshold. Until 30 September 2026, a covered provider was a GenAI producer with over 1,000,000 monthly visitors or users. It is now anyone that creates, codes or otherwise produces a GenAI system that is publicly accessible in California. A small image-generation startup that was out of scope in August is in scope now.

Second, AB 2713 replaced the platform text that AB 853 wrote. The 2025 version required platforms to detect provenance data compliant with widely adopted specifications that was embedded in or attached to content. The new version requires detection of any provenance data embedded in, attached to or otherwise associated with content, and a separate subsection says platforms need take no action on data that is not compliant or interoperable with those specifications. AB 2713 passed by majority vote as a non-urgency bill, so its text applies from 1 January 2027, the day the platform duties start.

In our searches we found no court order blocking any part of the act and no Attorney General guidance on it as of 12 November 2026.

Who is covered: providers, platforms, hosting platforms and device makers

The act assigns duties to four roles, and one company can hold several, such as a social app with an in-app image generator. Definitions come from section 22757.1 as amended by SB 1000; the operative dates come from sections 22757.3.1 to 22757.3.3 and 22757.6.

A few definitional details decide the edge cases:

• The 2,000,000 figure is unique monthly users. The definition does not limit the count to Californians.

• Broadband internet access and telecommunications services are excluded from the large online platform definition.

• Mass messaging means distribution to more than 100 users simultaneously. A one-to-one messenger is outside; broadcast channels and large groups may be inside.

• Content must reach users who did not create or help create it. In our reading, private storage where only the uploader sees a file is a weaker fit than a public sharing link.

• Exemptions are narrow. Before SB 1000, the act excluded services offering exclusively non-user-generated video game, television, streaming, movie or interactive experiences. SB 1000 cut that to exclusively non-user-generated videogames, and added a carve-out until 1 January 2029 for GenAI systems designed primarily to function as assistive technology.

What large online platforms must do from 1 January 2027

Section 22757.3.1, as rewritten by AB 2713, has five operative parts.

1. Detect. The platform must detect whether any provenance data is embedded into, attached to or otherwise associated with content it distributes. Provenance data means information about a piece of content's origin and modification history, in a format compliant or interoperable with widely adopted specifications from an established standards-setting body.

2. Display. The platform must provide a user interface that reliably indicates whether system provenance data or a digital signature identifies content as generated or substantially altered by a GenAI system, or captured by a capture device. The interface must clearly and conspicuously show, at a minimum, whether provenance data is present, the name of the GenAI system or capture device that created or substantially altered the content, and whether digital signatures are present.

3. Allow inspection. Users must be able to inspect system provenance data easily. The platform can show it in its own interface, link to a website or app that displays it (including one run by a third party), or let users download the provenance data, subject to federal copyright law, in a format that cannot easily be attached to unrelated content.

4. Do not strip. To the extent technically feasible, the platform must not knowingly strip system provenance data or digital signatures from content uploaded to, distributed on or downloaded from the platform.

5. Limits. The platform need not maintain, display or allow download of personal information, and need not act on provenance data that is not compliant or interoperable with widely adopted specifications.

System provenance data is the narrower category behind the display and stripping duties. It is provenance data not reasonably capable of being associated with a particular user that describes the device, system or service that generated the content, or relates to content authenticity. That definition is the legal basis for keeping identity data out of your provenance interface.

The 2025 and 2026 versions of the section compare as follows.

Our reading of otherwise associated with is that it reaches manifests linked to content rather than embedded in it, such as a manifest held in a remote repository and found through a watermark or fingerprint. The statute does not say so in those words, so treat it as an engineering assumption to confirm with counsel.

Latent disclosures, C2PA and widely adopted specifications

The act never names a standard. For image, video and audio, the leading candidate is the Content Credentials specification from the Coalition for Content Provenance and Authenticity (C2PA), a Joint Development Foundation project. The newest version listed on the C2PA specifications site is 2.5.

The parts of C2PA that map onto the law:

• Manifest and claim. A C2PA manifest holds assertions about an asset and a signed claim. The signer's certificate identifies who attests to the data, which matches the act's definition of a digital signature: a cryptography-based method that identifies the user or entity attesting to the signed information.

• Hard binding. One or more cryptographic hashes tie the manifest to the asset, so a validator can tell whether content changed after signing.

• Soft binding. A content identifier that is either a fingerprint or an invisible watermark, used to match content even when its bytes differ. A credential with a soft binding that lets it be found in a manifest repository is what C2PA calls a Durable Content Credential.

• Trust lists. C2PA maintains one list of trust anchors for claim signers and a separate list for time-stamping authorities. Your validator should check against both.

On the provider side, section 22757.3 requires a latent disclosure, to the extent technically feasible, in image, video and audio content that a GenAI system creates or alters, other than by minor modification. It must convey, directly or through a link to a permanent website, the provider's name, the GenAI system's name and version, the time and date of creation or alteration, a unique identifier, and whether the system created or altered the content. From 1 January 2029 it must also say whether the system is designed primarily to function as assistive technology. The disclosure must be permanent or extraordinarily difficult to remove or tamper with, compatible with the provider's verification tool, and compliant or interoperable with widely recognised industry standards. A common design puts the descriptive fields in a signed C2PA manifest and adds an invisible watermark as the soft binding, because metadata alone rarely survives re-encoding.

The act lists eight minor modifications that do not count as alteration: changes to brightness, contrast or colour; sharpening; saturating; file resizing; scaling; cropping; file format conversion; and audio denoising or background-noise removal.

What engineering teams build

Thumbnailing, transcoding into adaptive bitrate ladders and EXIF scrubbing typically remove embedded provenance data. The January work is about reversing that default for one class of metadata while still removing personal data.

Ingest detection. On upload, parse embedded manifests in supported containers, check for sidecar or remote manifest references and, if you adopt soft-binding lookups, query watermark decoders or fingerprint services against manifest repositories. The open-source c2pa Rust crate from the Content Authenticity Initiative, licensed under both MIT and Apache 2.0, creates, signs, embeds, parses and validates manifests and exposes a C API. It is still a 0.x beta whose breaking changes ship on a release train roughly every two months, so pin versions.

Validation. Verify the claim signature, check the certificate chain against the C2PA trust list, recompute the hard-binding hash and record a validation state per asset. A manifest whose hash no longer matches the content should not be presented like a valid one.

Preservation. Store the original manifest store at ingest, keyed to the asset, before any transform runs. For derived renditions you can re-embed the manifest where the format allows, serve it as a linked sidecar, or have your pipeline act as a claim generator that signs a new manifest listing the original as an ingredient. The stripping ban now covers downloads, so the download path needs the same treatment as playback.

Display and inspection. Add a label and an information panel that answer the three mandated questions: is provenance data present, which GenAI system or capture device created or substantially altered the content, and are digital signatures present. C2PA publishes user experience guidance that is a sensible starting point for wording and icons. For inspection, show the full system provenance data, link to a verifier, or offer a download; a signed manifest with its hard binding intact fits the unrelated-content test far better than a loose JSON export.

Privacy filter. Separate system provenance data from identity assertions and other personal information before display. Manifests can carry identity assertions under the Creator Assertions Working Group (CAWG) specification, which the CAI library supports, and AB 2713 says you have no duty to show personal information.

Evidence. Log detection coverage, validation outcomes and stripping defects per day. With a per-day penalty, being able to show when a defect started and ended has direct dollar value.

A per-asset record might look like this:

            {
  "asset_id": "vid_8f2c91",
  "provenance_detected": true,
  "source": "embedded_manifest",
  "validation_state": "valid",
  "signer": "Example GenAI Inc.",
  "generator": "ExampleVideo 3.1",
  "origin_type": "genai_created",
  "signatures_present": true,
  "manifest_store_uri": "s3://provenance-store/vid_8f2c91.c2pa",
  "personal_info_redacted": true,
  "checked_at": "2027-01-01T00:04:12Z"
}
          

If you also ship generative features, you are a covered provider for them. Embed latent disclosures and offer a free disclosure verification tool, your own or a compliant third-party one, that accepts uploads and URLs, can be called through an API, outputs any system provenance data it finds and returns personal information only with the person's express consent. Collect feedback on the tool's efficacy. When you grant a licence for the system, tell the licensee about its obligations. If you learn that an identifiable licensee modified it so it no longer complies, you must terminate the licence or notify the licensee within 72 hours; the licensee then has 96 hours to fix it or stop using it and report back, and if it does not, you report it to the Attorney General.

Model hubs hosting downloadable weights carry a narrower duty: do not knowingly offer a GenAI system that fails to place the required disclosures. An intake attestation plus a takedown route for reported models is a sensible minimum.

Does this apply to you? A decision table

Penalties and a worked example

Section 22757.4 sets a civil penalty of $5,000 per violation, recovered in a civil action by the Attorney General, a city attorney or a county counsel, with reasonable attorney's costs and fees for a prevailing plaintiff. Each day a covered provider, large online platform or capture device manufacturer is in violation is a discrete violation. Enforcement sits with those public officials; the chapter creates no private right of action. A separate penalty of $50,000 per violation, also counted per day, applies until 1 January 2029 to providers that falsely claim a GenAI system is designed primarily to function as assistive technology. SB 1000 also bars actions filed before 30 September 2026 from continuing over conduct that is lawful under the amended text.

Worked example. A US video-sharing app has 2.6 million unique monthly users through 2026 and ships its provenance interface on 15 February 2027. It is out of compliance from 1 January to 14 February, which is 31 + 14 = 45 days.

• Counted as one continuing violation: 45 × $5,000 = $225,000.

• Counted as four duties missed each day (detect, display, inspect, no-strip): 4 × 45 × $5,000 = $900,000.

• Counted per affected upload, which the text does not clearly rule out, exposure would scale with volume and could pass both figures.

How courts will count violations is untested, but a delay of weeks can reach six figures before legal costs.

Threshold check. Suppose the same app had 1.7 million unique users in January 2026 and 2.4 million by October. The definition covers a platform that exceeded 2,000,000 unique monthly users during the preceding 12 months without saying whether one month over the line is enough or an average is meant. Engineering should plan as if one month counts until counsel says otherwise.

How this fits with the EU AI Act and New York

In the EU, Article 50(2) of the AI Act requires providers of generative AI systems to mark outputs in a machine-readable format so they are detectable as artificially generated, and the regulation applies from 2 August 2026. Our EU AI Act Article 50 content marking guide covers the transition period for systems already on the market. California goes further by putting duties on distribution platforms, not only on generators.

New York's RAISE Act, which also starts on 1 January 2027, is a different kind of law. It sets safety framework, transparency and critical safety incident reporting rules for frontier AI developers, not content labels. Our New York RAISE Act explainer covers it, and a company can be subject to both, for example a large AI lab that also runs a consumer video app. For California's separate chatbot rules, see AI chatbot laws in 2027, and for every US and UK date in one place, our 2027 compliance calendar for software teams.

Checklist for the 50 days to 1 January 2027

1. Confirm your roles: large online platform, covered provider, hosting platform, or several. Pull 12 months of unique monthly user counts.

1. Inventory every media path that touches uploads: transcoders, thumbnailers, CDN image transforms, EXIF scrubbers and download endpoints.

1. Choose a C2PA library and version, and decide whether soft-binding lookups ship at launch or later.

1. Build ingest detection and validation, and store the original manifest before any transform runs.

1. Fix stripping in transforms and downloads, or move to linked manifests.

1. Ship the label and panel that answer the three required questions, plus an inspect or link path.

1. Filter personal information and identity assertions out of what you display.

1. Add per-day monitoring and an incident log for provenance defects.

1. If you ship GenAI features, check your latent disclosure fields and verification tool against the SB 1000 text.

1. Take three interpretation calls to legal: threshold counting, the reach of otherwise associated with, and how violations are counted.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI and machine learning services team builds C2PA ingest and validation, provenance-preserving transcoding and download paths, and the label and inspection interface, while the legal reading stays with your counsel. Talk to us at /contact-us/.

Last updated: 12 November 2026.

References

1. Business and Professions Code, Division 8, Chapter 25, AI Transparency Act, sections 22757 to 22757.6 (text as of 27 September 2026) — California Legislative Information

1. AB-2713 California AI Transparency Act: system provenance data, Chapter 856, Statutes of 2026 — California Legislative Information

1. California's nation-leading AI framework just got stronger, 30 September 2026 — Office of Governor Gavin Newsom

1. SB-1000 California AI Transparency Act, Chapter 861, Statutes of 2026 — California Legislative Information

1. AB-853 California AI Transparency Act, Chapter 674, Statutes of 2025 — California Legislative Information

1. Governor Newsom signs bills to further strengthen California's leadership in protecting children online, 13 October 2025 — Office of Governor Gavin Newsom

1. AB-2713 bill status — California Legislative Information

1. C2PA Specifications, version 2.5 index — Coalition for Content Provenance and Authenticity

1. Content Credentials: C2PA Technical Specification 2.5 — Coalition for Content Provenance and Authenticity

1. CAI Rust library (c2pa-rs) README — Content Authenticity Initiative on GitHub

1. Article 50: Transparency obligations for providers and deployers of certain AI systems — European Commission AI Act Service Desk

1. Article 113: Entry into force and application — European Commission AI Act Service Desk

1. DIGIT: AI Safety and the RAISE Act — New York State Department of Financial Services