Article 50 timeline: dates that bind
Four dates bind Article 50: 20 Jul 2026 guidelines, 2 Aug applies, 2 Dec provider marking grace, 2 Feb 2027 watermark-detection interoperability.
Article 50 of the EU AI Act is a calendar with four load-bearing dates. Put them in the production diary: 20 July 2026 for the Commission guidelines, 2 August 2026 for application, 2 December 2026 for the provider marking grace, 2 February 2027 for watermark-detection interoperability. This page is that diary. It is not another tour of the four transparency duties.
The Commission's FAQ on transparency obligations under Article 50 (last updated 24 July 2026) is the primary source for when the article starts to apply and for the limited grace on marking. Later dates for watermark-detection interoperability sit in the Code of Practice and in counsel notes published after the guidelines landed. Treat those notes as dated reading, not as a substitute for the FAQ.
The four dates on one calendar
Print this table. The rest of the page only explains why each row is a different kind of deadline.
| Date | What binds | Who it actually moves |
|---|---|---|
| 20 July 2026 | Commission adopts the final Guidelines on Article 50 transparency | Everyone who needed a reading of the article before it applied |
| 2 August 2026 | Article 50 applies | Providers and deployers. Deployer labelling of deepfakes has no grace |
| 2 December 2026 | Provider machine-readable marking for systems already on the market before 2 August | Providers of those systems. Not a holiday for publishers |
| 2 February 2027 | Watermark-detection interoperability backstop | Providers who owe a detection mechanism that other detectors can query |
Bird & Bird's first impressions of the final guidelines (20 July 2026, published as a 2026 insight) record that the Commission adopted the 51-page text less than two weeks before application. Paul, Weiss summarised the same package on 4 August 2026, including the 2 February 2027 interoperability date. Morgan Lewis, writing on 12 August 2026, confirmed that the Digital Omnibus deferred only specific marking work, not the bulk of Article 50.
Those four sources disagree on tone. They do not disagree on the first three dates. The fourth date is in the Code of Practice / counsel notes rather than in the FAQ's "when does Article 50 start" answer. Diarise it anyway. Detection that cannot talk to another detector is the thing the 2 February row is for.
A fifth date sits around the same week and is easy to confuse with the legal calendar: Meta and Google signing the Code of Practice on transparency of AI-generated content in late July 2026. That is a voluntary code, assessed as adequate by the Commission and the AI Board. Signing it is not a fifth legal deadline. It is a signal that large providers treated 2 December as an engineering date.
What 2 August switched on, and what 20 July was
20 July is not application. It is the day the Commission published the reading of Article 50 that market surveillance authorities can be expected to follow. The guidelines are non-binding; only the Court of Justice can give an authoritative interpretation. In practice, national authorities and, where Article 75(1) gives it exclusive competence, the AI Office will work from that text.
2 August is application. From that date, the FAQ is blunt: providers and deployers of AI systems must comply with the transparency obligations laid down in Article 50. A limited grace is envisaged only for AI systems placed on the market before 2 August 2026 and only as regards the marking and detection obligation for AI-generated content (Article 50(2)). Providers of those systems must comply with those marking obligations from 2 December 2026.
Two operational facts fall out of that sentence.
First, if you publish a photoreal talking head on 3 August, the deployer-facing label is already due. The tool you used may still be inside its marking window. Those are different clocks. The marking grace period and its limits is the ledger of which duties had no runway. This page only needs you to stop putting 2 December on a creator calendar as if it postponed the overlay.
Second, 20 July does not give you a two-week pardon. Content published between 20 July and 2 August is pre-application. Content published on or after 2 August is not. The guidelines exist so you could prepare. They do not move the start date.
The Cloud Security Alliance research note of 29 July 2026 is useful here for one reason only: it records that Article 50 was left out of the Digital Omnibus deferral that pushed Annex III high-risk systems to December 2027. If someone on the team still has "AI Act delayed" in a slide, that slide is about a different article. Article 50 was not in that delay.
The two later dates that are not a publisher holiday
2 December 2026 is a provider date. It attaches to a system that was already on the EU market before 2 August. It does not attach to a brand, an agency, a TikTok login, or a folder of unreleased takes. If the generator you prompt was placed on the market before 2 August, its provider has until 2 December to bring machine-readable marking and detection into line. If the generator was placed on the market on or after 2 August, marking was due from day one.
Morgan Lewis (12 August 2026) is explicit that the four-month transition does not postpone deployer duties. Emotion recognition notices, biometric-categorisation notices, deepfake labels, and public-interest text labels have applied since 2 August 2026.
2 February 2027 is also a provider date. Paul, Weiss (4 August 2026) describes the Code of Practice requirement: by that day, providers must make their watermark-detection mechanisms interoperable, so content can be verified without running each vendor's detector in turn. The Code gives examples — an industry-standard query-routing method, an embedded readable signpost, or a shared provider-agnostic solution. That is infrastructure. It is not a second chance to skip an on-screen line.
If you only prompt a hosted generator and then publish, you are not the addressee of either later date. You still cannot hide behind a missing watermark. The FAQ is explicit that deployers cannot simply rely on the machine-readable marking embedded by the provider to fulfil the disclosure a person has to perceive.
The Article 50 disclosure template is the copy and the three EU icons. Keep this page for when those artefacts are due. Keep Article 50 for the name of the rule. Keep the calendar for the dates.
How to diarise a production week against these dates
You are reading this after 2 August 2026. Three of the four dates are already in the past. That does not make the calendar decorative. It tells you which clock you are currently on.
This week, and every week until 2 December 2026. Assume deployer labelling is live. Assume some generators you use are still inside the provider marking window. Export the file, burn a perceivable disclosure on anything that would look like a real person, place, or event, and do not wait for a provenance tick to appear in the download. When a vendor switches marking on mid-window, the bytes may start carrying C2PA or a watermark that were absent in July. That is the provider hitting its own point on the 2 December runway. It does not retire your overlay.
From 2 December 2026. Treat missing machine-readable marking on a system that was already on the market as a vendor defect, not as a legal vacuum. New systems should already have been marking. Old systems have now run out of grace. You still label at first exposure. You now have a reason to ask a vendor whether detection actually works.
From 2 February 2027. Ask whether a detector that is not the vendor's own detector can read the mark. Interoperability is the row. A green tick that only the original host can verify is the thing this date was meant to end.
The generation / publication cut. The FAQ (24 July 2026) says content generated prior to 2 August 2026 does not need to be labelled retroactively, and the Commission encourages voluntary labelling of older published work. Paul, Weiss (4 August 2026) adds a distinction the FAQ compresses: content generated — or, in the case of Article 50(4), published — prior to 2 August 2026 does not need to be retroactively labelled. Bird & Bird's reading of the final guidelines treats the generation date as the cut for image, audio, and video deepfakes, and the publication date as the cut for public-interest text.
The working rule for a producer sitting on a drive of unreleased takes is the conservative one: if you generate it before 2 August and publish it after, treat the publication as covered. If you generated and published before 2 August, the article is not asking you to walk the archive. Do not spend December relabelling 2025 posts unless a client or a platform asks. Do spend the hour labelling anything you are about to post.
A clip from Veo 3.1 made in July and posted in September is a September problem. The model page is not a date. The upload is. If you need a studio seat to burn the line into the file rather than hoping a description field survives a download, sign up and treat the overlay as part of the edit.
Personal, non-professional use stays outside the Act. Regular economic benefit — a paid creator, a brand, an agency — is professional use. That sorting is not a date. It is a status. The calendar above assumes you already know you are in the professional column.
FAQ
Does the 2 December 2026 grace cover a missing deepfake label?
No. The FAQ limits the grace to Article 50(2) marking and detection for systems placed on the market before 2 August 2026. Deployer labelling of deepfakes applied on 2 August with no runway. A generator that is still catching up on watermarks does not postpone the visible or audible disclosure.
If I generated a clip in July 2026, when does the clock start?
The FAQ says content generated prior to 2 August 2026 does not need to be labelled retroactively. If you also published it before that date, you are not being asked to go back. If you are publishing it now, treat the post as covered and put the disclosure on the file. Public-interest text is the cleaner case: the publication date is the one that counts.
Who owns the 2 February 2027 date?
Providers who owe a watermark-detection mechanism. The Code of Practice, as summarised by Paul, Weiss on 4 August 2026, wants those mechanisms interoperable by 2 February 2027 so a verifier does not have to run each vendor's detector in turn. A creator prompting a hosted model does not implement that interoperability. They still cannot point at a missing detector and skip the on-screen line.