Say it was made with AI

There is exactly one genuine grace period, and it is narrower than people hope.

• Generative systems placed on the market before 2 August 2026 must comply with the Article 50(2) marking obligation only from 2 December 2026. That is a provider-side extension on machine-readable marking for tools that already existed.
• It does not extend your own duty. The deployer obligation in 50(4) has no such grace period. If you publish a deep fake today, the disclosure duty applies today.

There is also a voluntary layer sitting next to the law. The Code of Practice on Transparency of AI-generated Content, published 10 June 2026, had roughly 190 signatory organisations by late July 2026. Voluntary means voluntary. Signing it is a signal, not a defence, and not signing it is not a violation.

So the working position for tonight is this. The obligation on you is already running, the obligation on your tools finishes phasing in on 2 December 2026, and by the time anyone gets round to enforcing either one you will already have published the thing.

The word carries baggage from politics, and the baggage causes people to assume it does not apply to them.

The 50(4) duty attaches to content that has been artificially generated or manipulated in a way that resembles real people, places or events. A synthetic presenter reading your product update is inside that description. So is a voice clone of your CEO reading a script the CEO never read, even if the CEO approved it. Consent from the person depicted is not the same thing as disclosure to the viewer, and Article 50(4) is about the viewer.

What is outside it is the ordinary invisible use of AI in production. Cutting silence, cleaning up audio, generating captions, colour work. Nothing there resembles a real person doing something they did not do, so nothing there triggers a deep fake disclosure. That distinction, generation of apparent reality versus assistance with real footage, is the line every platform policy in the next section is also trying to draw, with varying success.

YouTube runs a label reading "Captured with a camera", derived from C2PA data. Its own help page tells you what the label's absence means, and it is refreshingly blunt: if it is missing, that does not mean the content has modified audio or visuals.

The same documentation describes the attack that breaks the label's presence. Point a C2PA-enabled camera at a screen showing synthetic content. The camera did genuinely capture what was in front of it, so it signs a genuine capture credential. The result is fully synthetic footage carrying a valid, honestly issued "captured with a camera" claim.

This is not a bug anyone can patch. It is the boundary of what a cryptographic binding can say: the camera can attest to what reached its sensor, and it cannot attest to what was in the room. Every provenance system built on capture-time signing has this hole, and the reason to teach it is not cynicism. It is so that when a credential is produced as evidence in front of you, you ask the right next question.

If manifests can be stripped, the obvious complement is a mark that lives in the content rather than beside it. That is what a soft binding is, and SynthID is the largest deployed example.

It embeds a signal directly in the pixels or the audio samples, and it is designed to survive cropping, filters, frame-rate change and lossy compression. Google reported over 10 billion pieces of content watermarked as of May 2025.

Read the two layers together rather than as rivals. The hard binding gives you a cryptographically strong claim that is easy to remove. The soft binding gives you a weaker claim that is hard to remove. Durable Content Credentials are the attempt to combine metadata, watermark and fingerprint so that stripping one does not strip them all. Neither layer, alone or together, tells you whether the content is honest. They tell you how it was made, and only when they survive.

Most disclosure anxiety is spent on things nobody is asking you to disclose. From YouTube's own policy page, no disclosure is required for:

• Beauty filters. The thing people assume is the most obviously synthetic layer on the internet is explicitly out of scope.
• Colour adjustment. Grading, correction, look-up tables. Out.
• Caption creation. Everything session 4 was about generates no disclosure duty at all.
• Cloning your own voice for your own dubs. This is the one that stops the room. Your voice, your content, another language: no label required.

TikTok adds one of its own. Generic text-to-speech does not require disclosure when the voice is not a recognisable voice of a known individual. The pivot is not synthesis, it is identity. A synthetic narrator is fine. A synthetic narrator that sounds like a specific real person is not.

Notice how closely that tracks Article 50(4). The law and the platforms have converged, from different directions, on the same underlying question: does this content make a viewer believe a real person did something they did not do? Once you hold that question, most of the rules stop needing to be memorised.

Set LinkedIn against the other three and the difference is structural, not cosmetic.

Its duty is deception-scoped. Disclosure is required for synthetic media that misrepresents a real person or real events, with an explicit parody carve-out. It is not a blanket label-all obligation. That is the Oversight Board's diagnosis, implemented: the question is about misleading the viewer, not about which tool was used.

Its label is not a label. LinkedIn renders a C2PA provenance panel built from the manifest the upstream tool signed. It is display-only. It does not sign anything, does not assert anything of its own, and if no manifest survived the upload there is nothing to display.

Both halves are honest, and the second one has a sharp edge. A LinkedIn post with no provenance panel tells you nothing whatsoever. That is not a flaw in their implementation, it is the strippability problem from Part 2 showing up in a real product. Which is exactly why the deception-scoped duty has to exist next to it, and why it is doing more work than the panel is.

Three live instruments and one that keeps getting cited as though it were law.

• TAKE IT DOWN Act. The platform notice-and-removal duty went live on 19 May 2026. Removal as soon as possible and not later than 48 hours, plus reasonable efforts to find known identical copies. Enforced by the FTC. Its digital forgery definition explicitly covers content made using software, machine learning or artificial intelligence.
• California AI Transparency Act, SB 942 as amended by AB 853, operative 2 August 2026. Above 1,000,000 monthly users: a free public AI detection tool with an API, optional manifest disclosure, mandatory latent disclosure. Large online platforms above 2,000,000 unique monthly users pick up provenance duties from 1 January 2027. Penalty: $5,000 per violation, each day counting separately.
• The detail that matters for procurement. The California statute does not name C2PA. It says "widely adopted specifications adopted by an established standards-setting body". If a vendor tells you the law requires C2PA specifically, they are selling, not citing.
• NO FAKES Act is still only a bill. Introduced, not passed. Treat every article describing it as a rule as a warning about that publication.

Note the shape across both jurisdictions. Europe puts the duty on the publisher and the provider. California puts it on the very large platform and the very large developer. Neither of them lets the individual creator off, and neither of them will arrive as a notification in your editing software.