California AB 853: the law that makes platforms read provenance, not just write it

A field note from ChronoVerify, built and maintained by a veteran intelligence analyst. Published July 26, 2026

Nearly every AI transparency rule written so far points in the same direction: whoever generates the content has to mark it. That is what makes them hard to act on if you are not a generator. You can read the rule, agree with it, and still find that it asks nothing of you. California's AI Transparency Act, as amended by AB 853 and signed on October 13, 2025, points the other way. It obligates the party distributing content to detect the provenance data already inside it, and to say whether that content came out of a model or out of a camera. This is a companion to our EU AI Act Article 50 checklist, which covers the write-side duties in Europe. The subject here is the read side, in California.

Three tiers, three dates

AB 853 amended the California AI Transparency Act (originally SB 942) and extended the existing compliance date while adding obligations for groups the original act did not reach. The duties and the dates below are taken from the bill text.

A large online platform is defined in the statute as a public-facing social media platform, file-sharing platform, mass messaging platform, or stand-alone search engine that distributes content to users who did not create or collaborate in creating it, and that exceeded 2,000,000 unique monthly users during the preceding 12 months. A capture device is a device that can record photographs, audio, or video, including still and video cameras, phones with built-in cameras or microphones, and voice recorders. System provenance data is provenance data not reasonably capable of being associated with a particular user, carrying either information about the type of device, system, or service that generated the content, or information related to content authenticity.

The question in the second branch

Read the platform disclosure duty closely and it contains two different questions wearing one sentence. Was this content generated or substantially altered by a generative AI system? Or was it captured by a capture device?

The detection industry has spent several years on the first question. Classifiers estimate whether pixels look synthetic, they are in a permanent arms race with generators, and they degrade on recompressed and screenshotted files. Whatever you think of that approach, it is at least aimed at the right target.

The second question is not the same question inverted. "Did a camera produce this file" is not answered by failing to find evidence of a model. It is a provenance question, and provenance is checkable: a capture device can sign what it recorded, and a verifier can confirm that the signature is intact, that the content binding holds, and that the signer chains to a recognized root. That is a cryptographic fact rather than an estimate, which is why it can carry an audit trail.

This is what ChronoVerify already does on images. It reads any embedded C2PA Content Credential, validates it against the official C2PA and CAI trust lists, and fails closed: a credential only counts as validated when the signer chains to a recognized root, so a self-signed or unknown signer is reported as present but unverified rather than confirmed. The result is a typed verdict on a five-point scale, and for anything that has to be defended later, a signed, timestamped report that a third party can check offline.

The part that changes the arithmetic

A read-side duty is only useful if there is something to read, and the honest state of things today is that most files carry nothing. Marking is stripped by resizing, re-saving, screenshots, and most upload pipelines. That is exactly why the January 1, 2028 tier matters more than its date suggests: capture device manufacturers must embed latent disclosures by default for devices sold in California. Combined with the platform no-strip duty from January 1, 2027, the statute works on both ends of the pipe at once. It requires the supply of capture provenance to grow, and it requires the platforms in the middle to stop destroying it.

That is a different bet from the detection arms race. It assumes provenance will become common enough to be worth reading, and then obligates somebody to read it.

What a platform can do about it now

The duty does not land until 2027, which makes this a good time to find out what your own content actually carries. Three steps, in increasing order of effort.

1. Measure your current strip rate. Take a set of files that carried credentials before upload, pull them back down through your own CDN and transforms, and check whether the credential survived. Most pipelines strip; the number is usually a surprise. No signup is needed to check a file.
2. Check what arrives. Run inbound content through a validating check to see what fraction carries a credential at all, and what fraction of those credentials validate against a trusted root rather than merely parsing. The gap between "has a manifest" and "validates as trusted" is where most of the interesting cases live.
3. Keep a record of the checks. A detection duty is easier to evidence when each check leaves a durable artifact. A typed verdict logged against the file's SHA-256 does that; a signed, timestamped report does it in a form a third party can verify without trusting the issuer.

Honest limits

Statutory language invites overclaiming, so the boundaries are stated plainly.

• This is information about a tool, not legal advice. The duties described here bind platforms, hosting platforms, covered providers, and device manufacturers. Confirm your own status and obligations with qualified counsel, and read the statute rather than this summary.
• No tool discharges the duty. There is no AB 853 certification for a product, and ChronoVerify does not claim one. A verification layer supports the work; the obligation stays with the obligated party.
• This is not an AI-generation or deepfake detector. ChronoVerify reads and validates provenance and reads the machine-readable marking a file carries. It does not score whether an image was made by AI, and pixel forensics do not detect AI generation.
• Absence of provenance is neutral. A file with no credential is a file with no evidence either way, not a suspicious file. Given how routinely marking is stripped, this will be the common case for some time.
• Validation is about the signer, not the scene. A validated credential means the signature and content binding are intact and the signer chains to a recognized root. It does not establish that what the image depicts is true.
• Images today. The statute's definition of a capture device covers photographs, audio, and video. ChronoVerify validates images. Audio and video are not supported.
• A verdict is investigative triage with documented limits, not courtroom proof. SOC 2 Type II and ISO 27001 are not held today. A verdict should not be the sole basis for an automated decision about a person.

Sources and further reading: AB 853, California AI Transparency Act (bill text, California Legislative Information); C2PA Specifications index; C2PA FAQ; EU AI Act Article 50 checklist; ChronoVerify compliance page; Marking checker; Method and documented limits.