The AI transparency regimes, on one timeline
Five laws now tell operators to mark, label, disclose and remove. Not one of them says how to prove you did. Each page below reads the statute, then shows the evidence the duty leaves behind.
01Two of them start on the same day
On 2 August 2026, California's AI Transparency Act becomes operative and the EU AI Act's Article 50 begins to apply. The alignment was deliberate: California moved its own date to match Brussels, so a company serving both markets faces one deadline rather than two.
The duties are not the same, though, and treating them as one project is the first mistake. California binds a single class of duty holder, prescribes four data elements, requires a public detection tool, and runs a 96-hour clock on licence revocation. Article 50 splits its duties between provider and deployer, prescribes no data elements at all, and has no detection-tool duty. One evidence discipline can answer both. One implementation cannot.
02The five regimes
In the order they took effect. Each links to a full reading of the statute — the duties, the thresholds, the penalties, and the evidence question each one raises.
-
China — AI content labelling Measures
Two labels: one a person can see, one written into the file. Binds generators, dissemination platforms, app stores and publishing users — the only regime here that regulates the whole distribution chain. Actively enforced since November 2025.
-
South Korea — AI Framework Act
Advance notice that a service runs on AI, a label on generative output, and a heightened duty where output is hard to tell from reality. Binding now; the regulator has deferred enforcement, which is a policy rather than a provision.
-
India — IT Rules, synthetically generated information
Labelling and embedded provenance for synthetic media, user declarations on upload, and removal on notice against four separate clocks running from two hours to seven days. No fine attaches — an intermediary that fails simply loses its safe harbour.
-
California — AI Transparency Act, SB 942
A latent disclosure carrying four statutory elements, a manifest disclosure the user can opt into, a free public detection tool, and licence revocation within 96 hours of discovery. $5,000 per violation, and each day counts separately.
-
European Union — AI Act, Article 50
Machine-readable marking of synthetic output by the provider; interaction, deep-fake and public-interest text disclosure by the deployer. Not deferred by the Digital Omnibus, though the marking duty has a transitional to December for systems already on the market.
03What they have in common
Read side by side, these regimes disagree about almost everything. They disagree about who is bound — the model developer, the publisher, the distribution platform, the app store, the end user. They disagree about method: China prescribes a metadata schema down to the field names, while the EU prescribes an outcome and names no technology. They disagree wildly about consequences, from a flat cap worth about twenty thousand US dollars to penalties measured in percentages of worldwide turnover.
They agree on one thing, and it is the thing none of them writes down. Every duty here is an act you perform and later have to demonstrate you performed. Embedding a disclosure, showing a notice, applying a label, verifying a file, revoking a licence, removing content inside a window — each is an event that happens once, in your systems, and leaves no durable trace unless you deliberately make one.
Not one of these five instruments tells you how to keep that record. Two of them get close. Korea's decree requires an operator to retain the basis for its high-impact measures for five years, and names a person who must vouch that those documents are current and accurate. China's Measures require six months of logs identifying who received unlabelled output. Both are retention duties on the evidence itself — and neither says a word about whether that evidence may be edited in the interim.
The gap is the whole product. An application log is written by the system whose behaviour is in question, rotates on a retention schedule, and can be edited without trace. A sealed record is written once, hash-chained to the ones before it, signed, and anchored — so a third party who trusts neither you nor us can confirm that what you are showing them is what was recorded, on the day you say it was.
04What Planisphere does not do
The same boundary applies on every page here, in each regime's own vocabulary. Planisphere is an evidence layer. It sits beside your systems, never inside the act.
- It does not mark, label, watermark, embed disclosures, or write provenance metadata. Those acts are the duty, and they are yours.
- It does not host, distribute, detect or remove content.
- It does not classify your system, decide whether an exemption applies, or determine which statutory window a notice falls into. It records the ground you relied on, and when — it does not endorse it.
- It does not certify compliance or make a legal determination. No vendor can, in any of these jurisdictions. A record is evidence a third party can check; it is not a finding that you complied.
- It does seal the acts you performed, in order, at the moment you performed them, in a form you cannot quietly revise afterwards.
Start the evidence trail
The console carries a pack for each of these five regimes, with every statutory act listed and the exact call that seals it. Sealing a record takes about a minute, and the first thousand each month are free.