California's AI Transparency Act requires certain large generative-AI providers to disclose the origin of AI-generated or altered images, video and audio. It requires detection tools, visible-label options, embedded provenance data and safeguards for licensed systems. Provenance data identifies where and when digital media was created or changed. The California Legislature says the law became operative August 2, 2026, and covers qualifying systems with more than one million monthly California users or visitors under AB 853.
Table of Contents
- Which companies and users are covered?
- What disclosures must providers add?
- How does the detection tool work?
- What happens when systems are licensed?
- What changes for online platforms in 2027?
Which companies and users are covered?
The law applies to providers of publicly accessible generative-AI systems that exceed the california usage threshold. its requirements focus on the companies providing those systems, not ordinary residents using them.
The threshold is more than one million monthly visitors or users in California. Products and services offering exclusively non-user-generated entertainment experiences are excluded. For users, that distinction matters. A covered provider must supply the required detection and disclosure features, but the law does not impose a general visible-label requirement on every person who shares AI-generated media.
What disclosures must providers add?
covered providers must give users the option to place a visible disclosure on generated or altered media. The notice must be clear, conspicuous, suited to the medium and understandable to a reasonable person. where technically feasible, the visible disclosure must also be permanent or extraordinarily difficult to remove.
Because the requirement is an option for users, not every covered output will necessarily carry a visible label. Providers must also embed a machine-detectable latent disclosure in generated images, video and audio when doing so is technically feasible and reasonable. The California Legislature's SB 942 requirements specify data identifying the provider, system and version, creation or alteration date and time, and a unique identifier.
How does the detection tool work?
Each covered provider must offer a free tool that assesses whether media was created or altered by that provider's system. Users must be able to submit an image, video or audio file through an upload or URL, and the provider must make detection available through an API. The tool must reveal detected system-provenance information, but not personal-provenance information.
📨 Get Free California Guides Alerts
Free · No spam · Unsubscribe anytime
Providers generally cannot retain submitted content or personal information longer than necessary to operate the tool. A negative result is not proof that a person created the media without generative AI. The required tool assesses whether the item came from that particular provider's system, so media from another system may fall outside its detection function.
What happens when systems are licensed?
A provider that licenses its generative-AI system must use contracts to preserve the system's latent-disclosure capability. If the provider learns that a licensee disabled that capability, it must revoke the license within 96 hours, according to the Legislature's licensing provisions. After revocation, the licensee must stop using the system.
This rule prevents a provider from avoiding the disclosure requirement simply by distributing its technology through another business. Violations can carry a civil penalty of $5,000 per violation for each day it continues. The Attorney General, city attorneys and county counsel may enforce the law.
What changes for online platforms in 2027?
Beginning January 1, 2027, online platforms with more than two million monthly users must detect qualifying provenance data and expose it through an interface. When technically feasible, they also cannot knowingly remove that data, as provided in AB 853.
This extends transparency duties from media creation to distribution. It does not make provenance data a fact-check: the information may identify a system and creation details without establishing whether the media's message is true. When evaluating questionable media: If a required feature appears to be missing, preserve the original file, its URL, the date and the provider or platform name before reporting the issue.
- Look for a clear visible disclosure.
- Use the named provider's free tool when the provider is known.
- Submit the original file or its URL, if available.
- Do not treat a negative result as proof that no generative system was involved.
- Beginning January 1, 2027, check a covered platform's interface for detected provenance data.
You Might Also Like
- California Has Its Own Privacy Regulator. Here Is What It Does.
- San Jose City Council Districts and Who Represents You
- How the San Francisco Board of Supervisors Districts Work