Don’t forget the four-month watermarking extension expires on 2 December

The EU AI Act’s Article 50(2) marking obligation carries exactly one grace period, and it closes on 2 December 2026. If a generative system you already run was on the market before 2 August 2026, that is your deadline — not next August’s.

Who this is for

• Providers who put a generative or content-manipulating AI system into the EU market before 2 August 2026 and assumed the whole of Article 50 lands in August.
• Product and platform teams responsible for synthetic audio, image, video or text output, including anything that can produce a deepfake.
• Compliance and legal leads mapping AI Act milestones and trying to work out which ones have a runway and which do not.

What’s happening

Article 50 — the AI Act’s transparency tier — applies from 2 August 2026. It requires that AI-generated or manipulated content be marked in a machine-readable way and be detectable as artificial. For most of Article 50, that is the date and there is no runway.

The one exception sits in Article 50(2). Providers whose generative systems were already placed on the market before 2 August 2026 get a four-month extension for the marking-and-detection requirement specifically. Their date is 2 December 2026. The European Commission’s own Article 50 guidance confirms it: existing systems must meet the marking obligation by 2 December 2026, and content generated before 2 August 2026 does not have to be labelled retroactively.

So there are two clocks. New systems: 2 August. Systems already live: 2 December. If you assumed the whole article lands in August, you have quietly given yourself a deadline four months too early — or, worse, missed that the December date is a hard stop, not a soft encouragement. December is already a crowded month for the Act; the Article 5 prohibition additions bite in the same window.

Why it bites

Marking is not a wrapper you bolt on the afternoon before. To be compliant the mark has to be machine-readable, reasonably robust to ordinary editing, and interoperable enough that a downstream detector can actually read it. That means picking a scheme — C2PA content credentials, a SynthID-style watermark, embedded metadata — and wiring it into the generation path, not the publishing path. Retrofitting that into a system that has been in production for a year touches the model output, the storage layer and often a CDN that strips metadata by default.

The interoperability problem is the part that catches people. A watermark only counts if something can detect it. Two vendors each shipping their own proprietary mark, neither readable by the other, is close to no mark at all in regulatory terms. I set out the C2PA-versus-SynthID trade-off in a companion piece, and it is worth reading before you commit to an approach.

The one thing to do before 2 December

Make one list: every generative or content-manipulating AI system you had in production before 2 August 2026. For each, write down what marking it emits today and whether an independent detector can read it. Any blank row is in scope for the 2 December deadline and needs a marking scheme chosen and integrated now — four months is enough time only if you start from a known inventory. Do the inventory this week; the integration is the slow part.

If you want a second pair of eyes on which of your systems are in scope, and on which watermarking approach survives your own pipeline, that is the kind of question I take on advisory.

Advisory is scoped and fixed-fee — you will know the number before any work starts.

The August headline will have passed by the time this matters to you. The date that still counts for anything you already run is 2 December 2026.