DEV Community

RuleReddy
RuleReddy

Posted on

California's AI Transparency Act after SB 1000: no user threshold, and what developers now have to build

California quietly rewrote its AI Transparency Act on September 30, 2026. If you skimmed the original SB 942 and concluded "this only hits companies with a million monthly users," that conclusion is now out of date.

SB 1000 (Stats. 2026, Ch. 861) was passed as an urgency statute, approved by the Governor and filed on September 30, 2026, and took effect immediately. This post walks through what the current text of Business and Professions Code sections 22757 to 22757.6 asks of teams that ship generative image, video or audio features, and how to build it.

What SB 1000 changed

According to the Legislative Counsel's digest of SB 1000, the bill:

  • deleted the user threshold from the definition of "covered provider" (previously "over 1,000,000 monthly visitors or users"),
  • renamed the "AI detection tool" to a "disclosure verification tool",
  • deleted the requirement to offer users an optional manifest (visible) disclosure,
  • added a new required field to the latent disclosure: whether the GenAI system created or altered the content,
  • exempted GenAI systems designed to primarily function as assistive technology until January 1, 2029, and added a $50,000-per-violation penalty for falsely claiming that status.

The definition of "covered provider" in section 22757.1(e) now reads in full:

"a person that creates, codes, or otherwise produces a generative artificial intelligence system that is publicly accessible within the geographic boundaries of the state."

No size floor. If your GenAI product is publicly accessible in California, read on.

The chapter itself has been operative since August 2, 2026 (section 22757.6, as set by AB 853).

The two core duties for covered providers

1. A latent disclosure in every image, video and audio output

Section 22757.3(a) requires, "to the extent it is technically feasible," a latent disclosure in any image, video or audio content (or a combination) that your GenAI system created or altered, except by "minor modification." The disclosure must convey, directly or through a link to a permanent website:

  • (A) the name of the covered provider,
  • (B) the name and version information of the GenAI system,
  • (C) the time and date of creation or alteration,
  • (D) a unique identifier,
  • (E) whether the system created or altered the content,
  • (F) from January 1, 2029, whether the system is designed to primarily function as assistive technology.

It must also be "permanent or extraordinarily difficult to remove or tamper with," compatible with your disclosure verification tool, and "compliant or interoperable with widely recognized industry standards."

"Minor modification" is defined in 22757.1(l): brightness, contrast or color changes, sharpening, saturating, resizing, scaling, cropping, format conversion, and audio denoising. So a resize endpoint isn't in scope; an AI "generative fill" is.

Note that text is not on this list. The latent-disclosure duty covers image, video and audio.

2. A free disclosure verification tool, with an API

Section 22757.2 requires a tool, at no cost, that lets a user check whether content was created or altered by your system, outputs any system provenance data it finds, and does not output personal information unless the person it pertains to expressly opted in. It must be publicly accessible (with reasonable anti-abuse limits allowed), accept an upload or a URL, and support "an application programming interface, that allows a user to invoke the tool without visiting the covered provider's internet website."

You may instead point users to a compliant third-party tool, as long as it reads your latent disclosures and is "clearly and conspicuously accessible through the user interface" of your GenAI system (22757.2(e)). You must also collect feedback on the tool's efficacy and must not keep personal information from it beyond what is strictly necessary.

Building the latent disclosure with C2PA

C2PA is the obvious "widely recognized industry standard" candidate. Here's a manifest that carries every field in 22757.3(a)(1). The standard c2pa.actions assertion covers "created vs altered"; a small custom assertion (reverse-domain label) carries the rest explicitly so your verification tool can read it without guesswork.

{
  "claim_generator_info": [{ "name": "acme-imagegen", "version": "4.2.1" }],
  "assertions": [
    {
      "label": "c2pa.actions",
      "data": {
        "actions": [
          {
            "action": "c2pa.created",
            "digitalSourceType": "http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia"
          }
        ]
      }
    },
    {
      "label": "com.acme.ca-disclosure",
      "data": {
        "provider": "Acme AI, Inc.",
        "system": "acme-imagegen",
        "system_version": "4.2.1",
        "created_at": "2026-11-10T14:05:22Z",
        "uid": "6f1c2e0a-9b7d-4a7e-8f0e-2d3b1c5a9e44",
        "operation": "created",
        "info_url": "https://acme.example/provenance"
      }
    }
  ]
}
Enter fullscreen mode Exit fullscreen mode

For an edit (say, an AI inpaint on an uploaded photo) use c2pa.edited and "operation": "altered". Generate the manifest per output rather than hard-coding it:

import { randomUUID } from "node:crypto";
import { writeFile } from "node:fs/promises";
import { execFile } from "node:child_process";
import { promisify } from "node:util";
const run = promisify(execFile);

const SYSTEM = { name: "acme-imagegen", version: process.env.MODEL_VERSION };

export async function discloseAndSign(inPath, outPath, operation /* "created" | "altered" */) {
  const uid = randomUUID();
  const manifest = {
    claim_generator_info: [SYSTEM],
    assertions: [
      { label: "c2pa.actions", data: { actions: [{
        action: operation === "created" ? "c2pa.created" : "c2pa.edited",
        digitalSourceType: "http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia",
      }]}},
      { label: "com.acme.ca-disclosure", data: {
        provider: "Acme AI, Inc.",
        system: SYSTEM.name,
        system_version: SYSTEM.version,
        created_at: new Date().toISOString(),
        uid,
        operation,
        info_url: "https://acme.example/provenance",
      }},
    ],
  };
  const mPath = `${outPath}.manifest.json`;
  await writeFile(mPath, JSON.stringify(manifest));
  await run("c2patool", [inPath, "-m", mPath, "-o", outPath, "-f"]);
  return uid; // store it: your verification tool and audit log need it
}
Enter fullscreen mode Exit fullscreen mode

Use a production signer (KMS or a remote signer), not c2patool's built-in development certificate.

"Extraordinarily difficult to remove." Metadata alone gets stripped by screenshots and many re-encoders. The statute's standard is "to the extent it is technically feasible," but the practical pattern is two layers: a signed C2PA manifest plus an invisible watermark that encodes the uid, so your verification tool can look up the record even when the manifest is gone. Keep a server-side table keyed by uid for exactly that reason.

A minimal verification endpoint

The tool must accept uploads or URLs and expose an API. A sketch with Express:

import express from "express";
import fs from "node:fs";
import multer from "multer";
import { execFile } from "node:child_process";
import { promisify } from "node:util";
const run = promisify(execFile);
const upload = multer({ dest: "/tmp/verify", limits: { fileSize: 50e6 } });
const app = express();

app.post("/api/verify", upload.single("file"), async (req, res) => {
  try {
    const { stdout } = await run("c2patool", [req.file.path]);
    const report = JSON.parse(stdout);
    // Return SYSTEM provenance only. Strip anything user-identifying
    // (author, account IDs, location) unless that person opted in.
    res.json({ found: true, provenance: redactPersonal(report) });
  } catch {
    // No manifest: fall back to watermark decode -> uid -> server record
    res.json({ found: false, provenance: await lookupWatermark(req.file.path) });
  } finally {
    // 22757.2(c): don't retain personal info beyond what's strictly necessary
    await fs.promises.rm(req.file.path, { force: true });
  }
});
Enter fullscreen mode Exit fullscreen mode

Add a URL variant (fetch with size and timeout limits, and SSRF protection), rate limiting for the abuse carve-out in 22757.2(a)(4)(B), and a feedback form. Don't require sign-up: 22757.2(d) bars making access contingent on personal information beyond what's strictly necessary.

Licensing your model? You have a 72-hour clock

Section 22757.3(b) is easy to miss. When you license your GenAI system to a third party, you must notify them of their obligations. If you know an identifiable licensee modified it so it no longer complies, you have 72 hours to terminate their license or notify them. They then have 96 hours to fix it or stop and report back, and if they don't, you report them to the Attorney General. You aren't required to monitor licensees (22757.3(b)(5)), but you need a process when you find out:

# compliance/ca-licensee-noncompliance.yml
licensee: example-partner
discovered_at: 2026-11-10T16:00:00Z
notice_due_by: 2026-11-13T16:00:00Z   # +72h
action: notified        # or: terminated
licensee_report_due: 2026-11-17T16:00:00Z   # +96h from receipt (if notified at the deadline)
ag_report_filed: false
Enter fullscreen mode Exit fullscreen mode

Coming next: platforms and devices

AB 853 added two more layers. From January 1, 2027, "large online platforms" (over 2,000,000 unique monthly users) must detect standards-compliant provenance data, show users whether it's available, and not knowingly strip it (22757.3.1), and GenAI hosting platforms must not knowingly offer a system that doesn't place the required disclosures (22757.3.2). From January 1, 2028, capture device manufacturers must offer latent disclosures in devices first produced for sale on or after that date (22757.3.3). If your upload pipeline re-encodes images, test now whether it drops C2PA manifests.

Penalties

$5,000 per violation, with each day a discrete violation, enforced by the Attorney General, a city attorney or county counsel (22757.4).

If you also serve the EU

The EU AI Act's Article 50(2) separately requires providers to mark synthetic outputs in a machine-readable format. One signed C2PA pipeline with a watermark fallback can serve both, but the scopes differ: California's latent disclosure covers image, video and audio, while Article 50(2) also reaches text. Map each output surface against both.

Checklist

  • List every surface that creates or alters image, video or audio content reachable from California.
  • Confirm none of them is only a "minor modification" path.
  • Emit all five 22757.3(a)(1) fields per output, with a stored uid.
  • Add a second, more robust layer (watermark) where stripping is likely.
  • Ship a free verification tool with an API, no sign-up, and a feedback form, or link a compliant third-party tool in your UI.
  • Add the licensee notice to your license terms and a 72-hour response runbook.

This isn't legal advice. It's a developer's reading of the statute text linked above; check your specific situation with counsel.

If you also need to scope EU AI Act Article 50 duties for the same product, RuleReddy has a free Article 50 check that walks through the questions.

Sources

Top comments (0)