Two AI Transparency Laws, One Start Date
EU Article 50 and California's AB 853 both went operative on 2 August 2026. What each asks of a creator selling into both markets, and the gap neither closes.
Two AI transparency regimes went operative on 2 August 2026. Article 50 of the EU AI Act applied from that date, with the Commission's guidelines finalised thirteen days earlier on 20 July. California's AB 853 became operative the same day.
For a creator or a small studio selling into both markets, the useful framing isn't "two new laws to comply with." It's that the two regimes attach to different parties, and for most independent producers the direct duties are lighter than the coverage suggests. What changes is what your tools are obliged to give you, and what you can be expected to have done with it.
What each one attaches to
| EU AI Act, Article 50 | California AB 853 | |
|---|---|---|
| Operative from | 2 August 2026 | 2 August 2026 |
| Core mechanism | Transparency obligations, including machine-readable marking of AI-generated output | Detection tooling that surfaces provenance data |
| Who it primarily binds | Providers and deployers of certain AI systems | GenAI systems above 1 million monthly users |
| Later phase | Machine-readable marking for systems already on market: 2 December 2026 | Duties for large online platforms and GenAI hosting platforms: 1 January 2027 |
| Enforcement ceiling | Up to €15 million or 3% of worldwide turnover | Statutory, per the bill text |
Read those "who it binds" rows carefully, because they're the reason panic-level coverage was overstated for individual creators. AB 853's headline duty in its first phase lands on generative systems at scale, not on the people using them. Article 50 splits duties between the entity supplying a system and the entity deploying it, and an independent creator publishing a video is normally in the second category with a narrower set of obligations. We've mapped where a creator actually sits on that line, and it's worth establishing before you write a compliance policy for your own shop.
Article 50, from the producer's seat
The practical demands, in order of how often they bite:
Machine-readable marking of generated output. This is the 50(2) duty, and it falls on the supply side rather than on you. Your exposure is indirect: you need to know whether the tools in your chain are marking, and whether your own handling preserves the marking. The Commission's stated position is that no single technique currently meets the standard, so metadata and watermarking are expected together rather than as alternatives. That rules out treating C2PA credentials as a complete answer.
Visible disclosure to the viewer. Separate layer, separate audience. The Code of Practice finalised alongside the guidelines includes three official EU disclosure icons for this purpose.
The grandfather window. Systems on market before 2 August 2026 have until 2 December 2026 for the machine-readable duty. That's a real operational fact rather than a technicality: a tool you use today may not be marking output yet and may still be inside its window, which affects what your finished files contain right now. That's the whole subject of the December deadline nobody diarised.
The interaction with commercial disclosure. If the video is also a paid placement, the AI disclosure and the sponsorship disclosure are independent and both have to land. Stacking them badly is the most common failure, and how to sequence them is a craft problem more than a legal one.
AB 853, from the producer's seat
AB 853 does something the EU regime doesn't: it puts a tool in your hands.
Generative systems above one million monthly users must offer an AI detection tool, at no charge to the user, that surfaces provenance data for content. The intended audience is anyone who wants to check a file. That includes journalists and platforms. It also includes you, checking your own work.
That's a genuinely different posture from a disclosure mandate. Article 50 tells you what to attach. AB 853 creates infrastructure for verifying what's attached, which is the step that has been missing from every disclosure workflow since C2PA shipped. If you've ever wondered whether your export actually carries credentials rather than whether your editor claims to write them, this is the phase that answers it.
The second phase, from 1 January 2027, extends duties to large online platforms and to GenAI hosting platforms. That's the part with real consequences for distribution, and it's a little over four months out.
Where they agree, and the gap neither closes
They agree more than the coverage implies. Both are transparency regimes rather than restrictions on generation. Neither bans making AI content, neither requires permission to publish it, and neither distinguishes between good and bad synthetic media on aesthetic grounds. Both assume provenance data is the mechanism.
If you already run a disciplined pipeline where generations are signed, edits preserve the chain, exports carry the manifest, and the published video carries a visible disclosure, you're substantively compliant with both and the remaining work is documentation. The pipeline discipline is the hard part, and it's the part most people skip. Sign, strip, survive walks the chain step by step.
That's the overlap. The divergence is more interesting, and it will shape the next two years more than either statute does.
Both regimes answer "was this made by a machine, and can you tell." Neither answers "were you allowed to depict this person."
That's not an oversight, it's a scoping decision. Transparency and consent are different regulatory objects. But it means a fully compliant workflow, machine-readable marking intact, EU icon burned into the frame, provenance verifiable through a detection tool, can produce a video that discloses perfectly and still had no business existing.
The consent question is being answered somewhere else entirely:
- In contracts. SAG-AFTRA members approved the 2026 TV/Theatrical Contracts by 91.42% to 8.58%, on a ballot that closed 4 June 2026. Producers committed not to use a synthetic performer in a role a human would play unless it brings "significant additional value", and to a notice-and-bargaining schedule before they do. If they breach it, the union can arbitrate for damages that are not necessarily limited to what a human performer would have been paid.
- In platform policy. Meta shipped an Instagram feature in July that let users generate images from any public account's photos, auto-enrolled everyone, and pulled it three days later after SAG-AFTRA publicly urged users to opt out. No transparency law was implicated. The feature disclosed nothing incorrectly. It just had no consent story.
There's a second, narrower gap worth naming: neither regime reaches the re-upload. Provenance metadata is strippable, platform badges live in platform chrome and evaporate when a file is downloaded, and nothing in either statute obliges someone re-posting your video to restore what their re-encode destroyed. Burning the disclosure into the frame is the only part of the stack that survives that, which is a production decision rather than a legal one.
One posture that covers both
- Keep the generation and the edit in a pipeline where the chain is preserved rather than reconstructed.
- Burn a visible disclosure into the frame, using the official EU icon assets from the Commission, sized for phone playback.
- Use each platform's own AI disclosure toggle as an independent signal, and keep it consistent with what the file says. The profile-level versus post-level distinction matters here, since they're not interchangeable.
- Verify the export rather than trusting the project. The AB 853 detection tools exist for exactly this.
- Keep a rights record for anyone depicted. Neither law asks for it, and it's the thing that will actually cost you. Treat it with the same seriousness as usage rights on any other licensed asset.
FAQ
Do I personally have obligations under AB 853?
In its first phase, the duties fall on generative systems above one million monthly users, not on individual users of those systems. The second phase, from 1 January 2027, extends to large online platforms and GenAI hosting platforms. For most independent creators the immediate change is that better verification tooling exists, not that a new filing obligation appeared.
Does complying with the EU regime cover me in California?
Substantively, largely yes, because both run on provenance data and a disciplined pipeline satisfies the mechanism behind each. But they're separate regimes with separate scopes and separate enforcement, and the later phases diverge, so treat the overlap as convenient rather than as legal equivalence. If you're operating at a scale where either regime's direct duties attach, that's a question for a lawyer rather than a blog post.
Does the shared start date mean the two regimes were coordinated?
Not as far as anything on the record says, and it's worth not reading intent into a coincidence of dates. The practical effect is real either way: a producer selling into both markets faces one changeover rather than two, and the tooling ecosystem had a single date to build toward, which is part of why detection tools appeared in a cluster rather than trickling out.
What should I do about the consent gap in the meantime?
Treat depicting a real person as a licensing question, not a disclosure question. Get written permission for likeness use, keep it with the project files, and don't rely on the fact that your video is honestly labelled as AI-generated to cover you. The label says how it was made. It says nothing about whether you had the right to make it.