Yesterday I wrote about Meta demanding a face scan to prove a human is real. Today the same authenticity problem gets attacked from the opposite end: not "prove the human is real," but "mark when the content is not." On August 2, two of the world’s largest regulatory blocs began requiring exactly that, on the same day, by no accident.
In California, SB 942 — the AI Transparency Act — became operative. In the EU, the AI Act’s Article 50 transparency obligations began to apply. The California date was deliberately pushed to August 2 (via AB 853) specifically to align with the EU timeline. Provenance for synthetic media is now law on two continents at once, and the convergence is the story.
What each regime actually requires
California SB 942 targets generative-AI providers with more than one million monthly California users. Every image, video or audio file their system generates or substantially alters must carry a latent disclosure — a machine-readable provenance record embedded in the content itself, containing the provider’s name, the system name and version, a creation timestamp, and a unique identifier tying the artifact back to the generating system. Providers must also publish a free detection tool that anyone can use to check a file, and offer users an optional visible label. In practice "widely accepted industry standards" means the C2PA specification. Enforcement runs at $5,000 per violation, per day, via the state AG.
EU Article 50 is broader in scope and narrower in mechanism. Providers must apply a machine-readable mark to synthetic content and make it detectable; deployers must disclose deepfakes and AI-generated text on matters of public interest, and disclose when a user is interacting with an AI system at all. Crucially, these obligations apply to existing systems from day one — there is no grandfathering — with fines up to €15M or 3% of global turnover. The Commission adopted its Article 50 guidelines on July 20, giving everyone twelve days of notice.
Why provenance, not watermarking, is the right primitive — and why it is still fragile
It is worth being precise about the mechanism, because the coverage muddles it. There are two distinct techniques here, and the laws lean on both:
- Embedded provenance metadata (C2PA "Content Credentials"): a cryptographically-signed manifest attached to the file describing its origin. Strong, verifiable, and detailed — but it lives in the container, and the container is easy to strip. Screenshot the image, re-encode the video, upload to a platform that rewrites metadata, and the manifest is gone.
- Signal-domain watermarking: a perturbation baked into the pixels or audio samples themselves, designed to survive re-encoding. More robust to stripping, but weaker per-bit, degradable by aggressive transformation, and the subject of an active adversarial-removal literature.
Neither is a solved problem, and anyone who tells you watermarking "solves" AI content is selling something. The honest framing: provenance is a chain-of-custody tool, not a lie detector. It reliably tells you "this file came from system X at time T" when the mark is intact. It cannot tell you a mark was never there, because absence of a mark is indistinguishable from a stripped mark. That asymmetry is the entire practical limitation, and the regulations do not repeal it.
The convergence is the actual news
The individual laws matter, but the meta-event is that two large, independent jurisdictions landed on the same primitive — C2PA-style provenance — on the same day. That is how a de-facto global standard gets born: not by treaty, but by enough big markets independently mandating the same technical mechanism that compliance becomes cheaper than fragmentation. For anyone building generative tools, C2PA just stopped being optional infrastructure. If your image, video or audio pipeline does not emit signed Content Credentials, you now have a market-access problem in two of your largest markets, not merely a nice-to-have.
Note also the direction of the obligation. Both regimes put the primary marking duty on the providers of the generative systems — the frontier labs and their API surfaces — and a disclosure duty on deployers. That is the right allocation: the entity that generates the artifact is the only one positioned to sign it at the source. But it also means the compliance burden concentrates upstream, and everyone downstream inherits whatever the upstream mark can and cannot survive.
What I would do this week if I built with generative AI
- Audit your generation pipeline for C2PA emission. If you generate media programmatically — via API or self-hosted models — check whether Content Credentials are being written and preserved through your own processing. Your provider may sign at the source and your resize step may strip it.
- Separate "provider" from "deployer" duties for your org. If you run a covered generative service, you have marking obligations. If you merely publish AI content, your duty is disclosure — labeling deepfakes and AI-generated public-interest content. Know which hat you wear; many companies wear both.
- Do not treat detection tools as ground truth. An absent mark is not proof of human authorship, and a present mark is not proof of honesty about everything else in the file. Build your trust model around what provenance actually asserts, not what a headline implies it asserts.
- Watch the platform layer. The weakest link is the social and CMS platforms that rewrite or strip metadata on upload. Expect a second wave of pressure — and possibly regulation — pushing platforms to preserve credentials, because marking at generation is pointless if distribution erases it.
Authenticity is now being enforced from both ends of the pipe: verify the human on one side, label the machine on the other. Neither mechanism is airtight, and the seam between them — a stripped mark, a coerced selfie — is where the next few years of adversarial work will live. But August 2, 2026 is a real inflection point: the day "where did this come from" stopped being a courtesy and became, on two continents, the law.