California AI Transparency Act [What It Changes]

TL;DR

• The California AI Transparency Act is California's law that, as of August 2, 2026, forces large public generative-AI image, video, and audio systems used in the state to offer a free detector, a visible label option, and a hidden tag inside the file.

• Covered providers sit above 1,000,000 monthly visitors or users; the civil penalty is $5,000 per violation, and each day counts as its own violation, with no private right to sue.

• Large online platforms and GenAI hosting platforms pick up extra duties on January 1, 2027; capture-device makers follow on January 1, 2028 for devices first produced for sale in California.

• Agencies and other small operators are not usually "covered providers," but the label now travels with the creative they publish, so DAM, CMS, and ad-upload paths have to stop stripping it.

Key Takeaways

• SB 942 created the Act in 2024; AB 853 moved the start date from January 1, 2026 to August 2, 2026 and added the later platform and device waves.

• Manifest disclosure is a user option; latent disclosure is the default tag the generator must embed when technically feasible.

• Text-only generative output is outside this chapter; neighboring AB 2013 (training-data posts) and SB 53 (frontier-model safety) cover different jobs.

• The same August 2, 2026 calendar also starts EU AI Act transparency duties for synthetic media, so dual-market creative stacks should treat labels as one pipeline, not two.

• Pending SB 1000 had passed both houses and been enrolled as of August 30, 2026, but it is not the operative statute until it is chaptered.

The California AI Transparency Act is the California statute that requires a generative-AI system with more than 1,000,000 monthly visitors or users, publicly accessible in the state, to give anyone a free way to check whether a file came from that system, to offer a visible AI label, and to bury a durable provenance tag in the image, video, or audio itself.

If you run a 10-person marketing agency, a two-truck HVAC shop, or a solo clinic, you probably do not train those models. You still ship their output. The client reel with an AI voice-over, the listing photo an image model cleaned up, the hold-message clone on the clinic line — those files now leave the generator with a latent tag, and from January 1, 2027 a large platform is told not to strip compliant provenance data when the file is uploaded. That is an operations problem, not a philosophy problem. Your DAM, CMS, "download the master" email, and ad-network ingest can drop metadata the same way they already drop IPTC copyright fields. The shop that cannot show a client whether last week's asset still carries its tag will stall the same way a shop that cannot produce a signed work order stalls. This hub walks the statute, the dates, and the workflow change for teams that publish other people's generators, including shops already lining up marketing-agency automation around a CMS rather than a new legal department.

What happened, and why the date moved

California enacted the Act as SB 942 (Chapter 291), approved September 19, 2024, adding Chapter 25 of the Business and Professions Code. The enrolled 2024 text set the chapter to become operative on January 1, 2026 and already named the three covered-provider duties: a free public AI detection tool, a manifest-disclosure option, and a latent disclosure.

AB 853 (Chapter 674), approved October 13, 2025, amended Sections 22757.1, 22757.4, and 22757.6 and added 22757.3.1–22757.3.3. The amendment delayed the chapter-wide start to August 2, 2026, then staged large-online-platform and GenAI-hosting-platform duties for January 1, 2027 and capture-device-manufacturer duties for devices first produced for sale in California on or after January 1, 2028. The current codified chapter still opens with the 1,000,000-user covered-provider definition.

According to Morgan Lewis, the act applies to GenAI services that have more than 1 million monthly visitors or users and are publicly accessible in California. The 1,000,000-user threshold is still the covered-provider test. The same LawFlash notes that CAITA's requirements do not apply to AI-generated textual content, and that products providing exclusively non-user-generated video game, television, streaming, movie, or interactive experiences are carved out.

According to AI Compliance Atlas, Chapter 25 became operative on August 2, 2026 under Section 22757.6. That is the precise announcement date this hub uses as its freshness stamp: as of August 2, 2026, the covered-provider duties are live. The Atlas record, last verified August 13, 2026, also states that AB 853 (Stats. 2025, Ch. 674) is the amendment that moved the date while SB 942 was still dormant.

The date was not random. Article 113 of Regulation (EU) 2024/1689 says the AI Act "shall apply from 2 August 2026," with earlier slices for prohibitions and general-purpose models. The European Commission said that from 2 August 2026 it would begin enforcing the Act and that, on the same date, new transparency rules would require certain systems to tell users when they are interacting with AI and when content has been generated or altered by it. California lined its first wave up with that calendar rather than with the original January 1, 2026 start.

The three duties in plain language

Think of the Act as three bolts on the same file, not three separate products.

First, the detector. A covered provider must make an AI detection tool available at no cost. The tool must let a user upload a file or paste a URL, must support an API so nobody has to visit the vendor's website, must report any system provenance data it finds, and must not report personal provenance data. The provider may not collect or retain personal information from detector users except opt-in contact details for feedback, and it may not keep submitted content longer than needed to run the check. That privacy posture sits next to the California Consumer Privacy Act rules that already limit how businesses handle personal information, including the $25 million gross-annual-revenue trigger that decides whether many firms are CCPA "businesses" at all. The California Privacy Protection Agency is the privacy regulator for that statute; CAITA enforcement itself is a civil action by the Attorney General, a city attorney, or a county counsel.

Second, the visible label. The provider must offer the user an option to include a manifest disclosure in image, video, or audio (or a mix) created or altered by its system. The disclosure must identify the content as AI-generated, be clear and conspicuous for that medium, and be permanent or extraordinarily difficult to remove to the extent technically feasible. "Offer the option" is not the same as "stamp every export." A designer who unchecks the box still receives a file; the latent tag is the one that is supposed to remain.

Third, the hidden tag. The provider must include a latent disclosure in AI-generated image, video, or audio that, to the extent technically feasible and reasonable, conveys the covered provider's name, the system name and version, the time and date of creation or alteration, and a unique identifier — either in the file or through a link to a permanent website. The tag must be detectable by that provider's own detector, consistent with widely accepted industry standards, and durable. If the provider licenses the system, the contract must require the licensee to keep that capability. If the provider learns a licensee disabled it, the license must be revoked within 96 hours, and the licensee must stop using the system.

According to Morgan Lewis, noncompliance carries civil penalties of $5,000 per violation, with each day of violation counting as a discrete violation, and there is no private right of action. Covered providers face $5,000 per violation per day. State authorities may also seek injunctive relief and recover attorney fees.

Sources: California Business and Professions Code Chapter 25; AB 853; AI Compliance Atlas SB 942 record.

Who is in scope, and who is not

A covered provider is a person that creates, codes, or otherwise produces a generative AI system that has over 1,000,000 monthly visitors or users and is publicly accessible within California. Most agencies, HVAC shops, and clinics are users of those systems, not producers of them. The Act still changes their week because the files they publish now carry a tag their clients, platforms, and competitors can query.

AB 853 added three later roles. A "large online platform" is a public-facing social network, file-sharing service, mass-messaging platform, or stand-alone search engine that distributes other people's content and that exceeded 2,000,000 unique monthly users during the preceding 12 months. A "mass messaging platform" is a direct-messaging tool that can send to more than 100 users at once. A "GenAI hosting platform" is a site or app that lets a California resident download source code or model weights, paid or not. A "capture device" is anything that records photos, audio, or video, including phones; a capture-device manufacturer is the producer of that device for sale in the state, not a firm that only assembles it.

According to AI Compliance Atlas, capture-device-manufacturer duties apply to devices first produced for sale in California on or after January 1, 2028. Large platforms wait until January 1, 2027. Hosting platforms share that 2027 date: they may not knowingly make available a GenAI system that does not place the required disclosures.

Sources: AB 853 definitions in § 22757.1; codified Chapter 25.

SB 1000 is the live amendment fight, not current law. The bill status page shows an urgency measure by Sen. Becker, last amended August 21, 2026, with Assembly amendments concurred in August 27, 2026 (Ayes 39, Noes 0) and an enrolled date of August 30, 2026. Morgan Lewis, writing August 3, 2026, described the then-pending draft as a bill that could delete the 1,000,000-user threshold, scrap the manifest-disclosure option, tighten latent-disclosure rules, cut the license-revocation window from 96 hours to 72, and rename the detector a "disclosure verification tool." Until a chaptered bill replaces Chapter 25, the 1,000,000-user test and the three original duties remain the ones a workflow should implement.

How the hidden tag is supposed to survive the trip

"Consistent with widely accepted industry standards" is the statute's way of pointing at the provenance stack vendors already ship, not at a new California file format.

The Coalition for Content Provenance and Authenticity publishes Content Credentials, an open technical standard for origin and edit history. The current C2PA 2.3 specification describes assertions, cryptographically signed claims, and manifests that can be embedded in common file types so a later reader can see whether the credential still matches the bytes. Content Credentials describe the consumer-facing pin and state that the effort is a collaboration with 500+ companies, including Microsoft, Adobe, Intel, BBC, Truepic, Sony, Publicis Groupe, OpenAI, Google, Meta, and Amazon. The Content Authenticity Initiative is the related cross-industry effort aimed at the same transparency job.

Partnership on AI splits the toolbox into indirect disclosure (watermarks, signed metadata, hashing) and direct disclosure (labels people can see). That split maps cleanly onto CAITA's latent versus manifest pair. IPTC photo metadata remains the unsigned, widely deployed field set newsrooms already know; C2PA is the signed overlay. Agencies that already fight to keep IPTC copyright intact on social uploads are fighting the same class of bug they will fight for latent CAITA tags.

Vendors are not waiting for a California inspector. OpenAI said it began adding C2PA metadata to DALL·E 3 images in ChatGPT and the API, joined the C2PA Steering Committee, and opened an image-detection classifier to researchers. According to OpenAI, internal testing showed the classifier correctly identified ~98% of DALL·E 3 images, with less than ~0.5% of non-AI images tagged by mistake. OpenAI's DALL·E 3 detector hit about 98% on in-house tests. The same post says performance against other labs' models was lower, flagging about 5–10% of those images on an internal set, which is why a California "use our detector" duty is not a universal oracle.

Google DeepMind's SynthID watermarks images, video, audio, and text in Google's consumer products in a way people are not meant to notice, and it offers a Detector portal for files that carry that watermark. Meta said it would label images when it can detect C2PA and IPTC "AI generated" markers from other companies, and that it already applied "Imagined with AI" labels to photorealistic Meta AI images, using both visible markers and invisible watermarks. Those platform behaviors are why an agency's export path matters: a tag that dies in the CMS never reaches the network that would have shown it.

NIST's voluntary AI Risk Management Framework (released January 26, 2023) and the July 2024 Generative AI Profile (NIST AI 600-1) treat content provenance as a risk-management control, not a marketing badge. CAITA is the California statute that turns a slice of that control into a duty for large AV generators.

What a small team actually changes this week

You do not need to be a covered provider to feel the Act. You need a file path.

Map every place AI image, video, or audio enters the shop: the designer's Midjourney or Firefly export, the editor's voice clone, the salesperson's product demo, the clinic's translated hold message. For each path, write down whether the generator is above the 1,000,000-user line and whether the export still contains C2PA or equivalent latent data after it hits disk. Then walk the same file through the CMS, the DAM, the social scheduler, and the "send the client a ZIP" step. If any hop re-encodes without copying provenance, that hop is now a business risk even though the statute's $5,000 hammer lands on the provider, the large platform, or the device maker — not on a 10-person agency.

Treat the vendor's public detector as a QA gate, not a toy. Covered providers must expose an API. A production check can POST the final master and fail the job if the detector cannot see the tag the generator claimed to embed. Teams already routing creative through US Tech Automations workflows can add that detector call as a model-swap step on the same path that already moves a finished asset into the CMS.

Keep a manifest-disclosure default for client-facing photoreal work unless the contract says otherwise. The law makes the visible label an option the user can decline; brand and platform rules may still want it on. Store the choice next to the brief so an account lead can show why a given ad is unlabeled.

Do not strip tags "to make the file smaller." From January 1, 2027, large platforms are told not to knowingly strip compliant provenance data or digital signatures. An agency that uploads a cleaned file is training its own future self to fail that inspection. The same hygiene belongs in form-to-CRM intake when a client drops a folder of "finals" onto a web form: hash the upload, keep the original bytes, and run the detector before anyone "optimizes" the file.

Licensees of a covered system should read the 96-hour revocation clause. If you fine-tune or wrap a licensed generator and disable latent tags, the licensor is required to pull the license within 96 hours of learning that, and you must stop using the system. That is a contract and a runtime problem for any shop that ships a white-label creative tool.

A 10-person agency comparing project-management stacks should put "preserve provenance on export" on the same scorecard as time tracking. The small-business automation picture is the same: the win is a repeatable path, not a one-off legal memo. US Tech Automations can sit in the pipeline that already files the invoice and publishes the asset, so the label check is one more node rather than a new department.

Neighbor rules you should not mix up

CAITA is one of three California generative-AI transparency statutes that operators keep conflating.

AB 2013 (Chapter 817, approved September 28, 2024) is the training-data documentation law. On or before January 1, 2026, and before each later public release or substantial modification of a generative system released on or after January 1, 2022, a developer must post high-level dataset documentation on its website. That is a webpage duty, not a file-tag duty, and it does cover text systems.

SB 53 (Chapter 138, approved September 29, 2025) is the Transparency in Frontier Artificial Intelligence Act. It binds large frontier developers to publish a frontier AI framework and to report catastrophic-risk assessments. "Catastrophic risk" in that bill includes a foreseeable, material risk of death or serious injury to more than 50 people or more than $1,000,000,000 in property damage from specified model behaviors. That is a safety-reporting statute, not a watermark statute.

On the other side of the ocean, Regulation (EU) 2024/1689 is the AI Act. Article 50 requires providers of systems that generate synthetic audio, image, video, or text to mark outputs in a machine-readable format, and it requires deployers of deepfakes to disclose that the content was generated or manipulated, with a narrower rule for public-interest text. The Commission published transparency guidelines on 20 July 2026 ahead of the 2 August 2026 start, and its AI Act overview states that the transparency rules come into effect in August 2026. According to the European Commission, more than 180 organisations had signed the Code of Practice on transparency of AI-generated content by the enforcement date.

Sources: AB 2013; SB 53; Commission 2 August enforcement note; California AG CCPA page.

USTA analysis: two clocks you can recompute

This is a derivation from figures already cited, not a new statistic.

Input A, original operative date in SB 942 § 22757.6: January 1, 2026. Input B, amended operative date in AB 853 § 22757.6: August 2, 2026. 2026 is not a leap year, so January–July contain 31+28+31+30+31+30+31 = 212 days, and August 2 is day 214 of the year. Day 214 minus day 1 equals 213 days. AB 853 therefore moved the first covered-provider duties 213 days later than the 2024 statute had printed, lining them up with the EU AI Act's 2 August 2026 application date.

Input C, civil penalty in § 22757.4: $5,000 per violation. Input D, the same section: each day a covered provider, large online platform, or capture-device manufacturer is in violation is a discrete violation. For a single continuing miss that lasts 30 calendar days, 30 × $5,000 = $150,000 before injunctive relief and attorney fees. Thirty is the analysis window, not a number in the statute; $5,000 and daily stacking are. A reader can swap in 7 days ($35,000) or 90 days ($450,000) with the same arithmetic.

USTA analysis. Inputs: SB 942; AB 853; BPC Chapter 25.

Signal vs Speculation

Demonstrated fact, sourced above: Chapter 25 is operative as of August 2, 2026 for covered providers; the detector, manifest option, and latent tag are live duties; the penalty is $5,000 per violation with daily stacking for named entity types; there is no private right of action; text is outside this chapter; AB 853's 2027 and 2028 waves are in the code; SB 1000 was enrolled August 30, 2026 and is not the current chapter; EU Article 50 transparency duties apply from 2 August 2026.

Our read: if the 1,000,000-user floor survives, most agencies stay off the covered-provider list and the operational pain stays in file hygiene — keep the tag, log the detector result, stop re-encoding masters. If SB 1000 is signed in the form Morgan Lewis summarized, the floor disappears and any public GenAI system accessible in California could inherit detector and disclosure duties; that would pull white-label agency tools and small open-weight hosts into the same compliance set as the labs. We do not treat enrollment as enactment.

Our read: the 2027 large-platform wave is the date that will hit marketers even if they never build a model. Once Facebook-scale properties must surface provenance and must not strip it, unlabeled or tag-stripped creative becomes a distribution defect, the way a missing tracking pixel is a distribution defect today. Shops that wait until 2027 to preserve C2PA on export will be debugging live campaigns instead of a staging folder.

Our read: if X holds — if C2PA-class credentials remain the "widely accepted industry standard" the statute points at — then a single preserve-and-verify node covers California and the EU for image, video, and audio. If platforms keep dropping unsigned IPTC while keeping signed C2PA, agencies should prefer the signed path and should stop treating "save for web" as a harmless click.

Our read: SMBs that only generate text (captions, blogs, email) are not in CAITA's file-tag net, but they can still be in AB 2013's documentation net if they are developers, and they can still be in EU Article 50 if they publish unreviewed public-interest text into the Union. Do not use this hub as a "text is free" memo.

Glossary

• Covered provider. A person that creates, codes, or otherwise produces a generative AI system with over 1,000,000 monthly visitors or users that is publicly accessible in California.

• Manifest disclosure. A label a person can perceive that identifies content as AI-generated; under CAITA it is an option the user must be offered, not an automatic stamp on every export.

• Latent disclosure. Provenance data that is present but not obvious — typically metadata or a watermark — that names the provider, system version, timestamp, and a unique ID.

• System provenance data. Provenance that is not reasonably linkable to a particular user; the detector must output this and must not output personal provenance data.

• Large online platform. A social, file-sharing, mass-messaging, or search service that distributes other people's content and that exceeded 2,000,000 unique monthly users in the prior 12 months.

• Content Credentials / C2PA. The open standard for signed, tamper-evident origin and edit history that vendors currently treat as the industry provenance format.

• AI detection tool. The free, public, API-capable service a covered provider must ship so anyone can ask whether a file was made or altered by that provider's system.

Frequently asked questions

Does the California AI Transparency Act apply to a 10-person marketing agency?

Usually not as a covered provider, because that role requires producing a public GenAI system with more than 1,000,000 monthly visitors or users. The agency still has to keep labels on the files it publishes, because clients, platforms, and later 2027 platform rules will treat a stripped tag as a defect.

Is AI-written copy covered?

No. Morgan Lewis and the statutory definitions confine CAITA's detector and disclosure duties to image, video, and audio (or combinations of those). Text systems can still fall under AB 2013 documentation rules or under EU Article 50 if the work is published in the Union.

Can a customer sue if an ad is unlabeled?

No private right of action exists under Chapter 25. Enforcement is a civil action by the Attorney General, a city attorney, or a county counsel for $5,000 per violation, with each day a discrete violation for covered providers, large online platforms, and capture-device manufacturers.

What is the difference between the California AI Transparency Act and the EU AI Act?

CAITA is a California provenance-and-detector statute for large AV generators, with later platform and device waves. The EU AI Act is a Union-wide product-safety and transparency regulation; Article 50's marking and deepfake-disclosure rules apply from 2 August 2026 and also reach certain text. The calendars match; the duties do not.

When do social networks have to show the tag?

January 1, 2027, for large online platforms as defined — including a duty to detect compliant provenance data, show it in the UI, let users inspect it, and not knowingly strip it.

What should we do if a licensed model we wrap has its tags turned off?

The covered provider must revoke that license within 96 hours of learning the tags were disabled, and the licensee must stop using the system. Treat a disabled watermark as an incident, not a feature flag.

Where does US Tech Automations fit?

US Tech Automations is the workflow layer that can call a vendor detector API, refuse a publish step when provenance is missing, and keep the original bytes beside the optimized export — the same pattern already used to move forms into a CRM.

Prepare the stack as if the tag is part of the master file, because as of August 2, 2026 it is. Map the generators you actually use, keep originals, and put a detector check on the publish path — then wire that check into an agentic workflow so the label survives the trip from tool to client. The US Tech Automations homepage is the index for those workflow products if you are starting from a blank shop.