The December Deadline Nobody Diarised: Grandfathered Tools and Article 50

The assumption trap goes like this: a tool you've been using since before August 2026 has never shown you a watermark, never popped up a disclosure prompt, never added visible metadata to an export. Nothing about using it feels different after August 2 than it did in July. So the reasonable-sounding conclusion is that it's fine — compliant, sorted, nothing to diarise. That conclusion is wrong for a specific, easy-to-miss reason: some of the machinery that's supposed to make that tool's output technically detectable as AI-generated isn't required to exist yet. It has a second deadline, in December, and almost nobody who isn't reading EU Commission FAQs on purpose knows it's there.

This is a plain-language walkthrough of that second clock — not legal advice, and not a replacement for counsel if a campaign has real EU exposure. What follows is the map worth bringing to that conversation.

Two clocks, not one

Article 50 of the EU AI Act isn't a single obligation with a single deadline — Versely's earlier breakdown of what Article 50 covers goes through the full structure, including exactly what counts as a regulated deepfake. The piece that matters here is that two of its obligations sit on different people and, as it turns out, different timelines.

Article 50(4) is a deployer duty — it falls on whoever publishes realistic synthetic content, and the Commission's own FAQ confirms Article 50 has applied since August 2, 2026, with the deepfake-disclosure test turning on whether the content would falsely appear authentic or truthful to its audience — not on whether the person who made it intended to deceive anyone. That obligation has been binding, full stop, since that date.

Article 50(2) is a provider duty — it falls on whoever builds the generative AI system, and it requires marking outputs in a machine-readable format so the content is technically detectable as AI-generated after the fact. That's the one with the second clock.

What "grandfathered" actually means here

Per the Commission's FAQ on the transparency Code of Practice, the AI Omnibus proposal agreed by the EU co-legislators envisages a targeted grandfathering rule for the marking and detection obligations under Article 50(2), specifically for generative AI systems placed on the market or put into service before 2 August 2026. If adopted as agreed, that gives providers of those existing systems until 2 December 2026 to bring the marking and detection machinery into conformity.

Worth sitting with what that machinery actually is, because it's not a checkbox — it's engineering work. Machine-readable marking means something like embedded provenance metadata or a technical watermark that survives export and re-upload, detectable by tools built to look for it, not just a visible "AI-generated" caption a human reads. Building and rolling that out across an existing product, retroactively, for every generation type a platform supports, is exactly the kind of work that justifies negotiating a four-month runway instead of a same-day switch-on. That's also why "grandfathered" doesn't mean "exempt" — it means "has a deadline instead of already being required," and the deadline is real: 2 December 2026, contingent on the Omnibus proposal being adopted as currently agreed.

Why your tool's behavior may quietly change before then

Because the grandfathering window runs through December, and because different providers are building toward it on different internal schedules, the practical experience for anyone using these tools is that marking behavior can switch on mid-use rather than all at once on a known date. A watermark that wasn't there in August starts appearing in September exports. Metadata that used to be absent shows up in a file's properties. An in-app "this is AI-generated" toggle appears where there wasn't one the week before. None of that is a bug — it's a provider hitting its own point along the conformity runway.

That this is actively happening, not theoretical, shows up in the timing of who's signing the Code of Practice that operationalizes it. Meta confirmed it was signing the EU AI Act Code of Practice on Transparency of AI-Generated Content on July 28, 2026, and Google separately announced signing the same transparency Code around the same window. Two of the largest AI providers committing to the Code in the same few weeks the grandfathering clock started running is a fairly clear signal that the December date is being treated as a real engineering deadline industry-wide, not a soft target likely to slip quietly.

What doesn't wait for December

Here's the part worth being precise about, because it's the easiest thing to get backwards: the grandfathering clock belongs to the tool provider's marking obligation. It has nothing to do with your own disclosure duty as the person publishing content.

If a tool you've used since July still isn't watermarking its output in October, that tells you about the provider's conformity timeline — it tells you nothing about whether you personally still need to disclose realistic synthetic content to an EU audience under Article 50(4). You do, and you've needed to since August 2, regardless of what any given tool's export does or doesn't embed. Synthetic media disclosure and a proper AI content label are your obligation as the deployer, sitting entirely outside whether the tool that generated the asset has finished rolling out its own machine-readable marking. Waiting for a tool's watermark to switch on before you start disclosing is treating someone else's engineering deadline as if it were yours.

A practical response for the next few months

Since a tool's marking behavior can shift underneath a campaign that's already running, the sensible move is to stop depending on it entirely for your own records. Keep an independent account of what you generated, when, and with what — a paper trail that doesn't rely on any provider's export metadata catching up on its own schedule.

A lightweight version of that inside Versely: save generations to a dated project as a campaign runs, rather than reconstructing the timeline later from scattered files. A request like

"Save this week's generated avatar clips and voiceovers to a project called 'EU Campaign — August 2026 batch'"

keeps a running, timestamped record of exactly what was synthetic and when it was made — independent of whether the tool that made it has finished its own December conformity work, and genuinely useful if a review ever asks what shipped and when. It's a production habit, not a compliance product, but it's the kind of habit that makes a real compliance conversation shorter when it happens.

Before December 2

• Confirm which of your tools were placed on the market before August 2, 2026 — those are the ones riding the grandfathered clock, and their marking behavior is the one likely to change without notice between now and December.
• Don't treat the absence of a visible watermark as evidence you're covered. It may just mean that tool hasn't hit its own conformity milestone yet.
• Keep disclosing realistic synthetic content to EU audiences now, on the Article 50(4) timeline that's already binding — that duty was never grandfathered.
• Keep your own dated record of what you generated and when, so a provider's marking rollout timing doesn't become the only record of your campaign's synthetic content.

FAQ

Does the grandfathering rule mean I don't have to disclose AI content until December?

No. Grandfathering applies to the provider's technical marking obligation under Article 50(2). Your own disclosure duty as a deployer under Article 50(4) has applied since August 2, 2026, with no grandfathering mentioned for that obligation.

What's actually different between Article 50(2) and Article 50(4)?

50(2) requires AI providers to mark generated output in a machine-readable, technically-detectable format. 50(4) requires deployers publishing realistic synthetic content to disclose it to their audience. One is a build requirement on tool-makers; the other is a publishing requirement on whoever ships the content.

Is the December 2, 2026 date locked in?

It's the date set out in the Commission's FAQ on the AI Omnibus proposal as agreed by the EU co-legislators, contingent on that proposal being adopted as agreed. Treat it as the working date to plan against, and confirm final adoption status with counsel if a campaign has meaningful EU exposure.

What should I actually change before then?

Little in your own workflow — keep disclosing as you already should be. The change is in expectation: don't read a tool's silence on watermarking as a compliance signal, and don't be surprised if marking behavior changes mid-campaign as providers hit their own conformity dates through December.