What actually moved, and what did not
Article 50 has four operative limbs and only one of them moved. Chatbot disclosure under 50(1), emotion and biometric notice under 50(3), and deepfake and public-interest text disclosure under 50(4) all applied on 2 August 2026 exactly as drafted, and are live obligations now. The page on what has applied since 2 August 2026 works through each of them.
What moved is the Article 50(2) duty on providers of systems generating synthetic audio, image, video or text content to mark those outputs in a machine-readable format detectable as artificially generated or manipulated. The amending regulation gave that duty a transitional period ending 2 December 2026, and the transitional period reaches one population only: generative systems already placed on the EU market before 2 August 2026.
The framing matters because "the marking deadline moved to December" is heard as a general reprieve, and it is not one. For a product launched after the summer of 2026 the marking duty was never deferred at all.
The cut-off is placement, not launch
The test turns on placement on the market. Under Article 3(9) placing on the market means the first making available of a system on the Union market, and under Article 3(10) making available means supply for distribution or use in the course of a commercial activity, whether for payment or free of charge. Article 3(11) covers putting into service, which is supply for first use directly to a deployer or for the provider's own use for its intended purpose.
Three consequences fall out of that definition, and each one moves the answer for a real product.
Free does not mean unplaced. Making available free of charge is still making available. A free tier, a public beta or an open demo can constitute placement.
Building is not placing. The date you trained the model, finished the feature or announced it publicly is irrelevant. What counts is the first point at which it was supplied for use in the Union.
Union placement is what matters. A system available in other markets since 2024 but first made available to EU users in September 2026 was placed on the Union market in September 2026, and gets no transitional relief.
The asymmetry is worth stating plainly, because it runs against intuition. An older product carrying years of legacy generation pipeline gets four extra months. A new product built after the rules were known gets none. The regulation is treating retrofit cost as the thing worth relieving, not novelty.
Which side of the line are you on?
As of 2 August 2026 this is a question about the past, which makes it answerable with evidence rather than judgement. Work through it in this order.
- Does the system generate synthetic audio, image, video or text? If it does not, Article 50(2) does not reach it and the rest of this page is moot.
- Are you the provider of that system?The duty binds providers, not deployers. If you call someone else's generative API and publish the output, the marking duty sits with them, though a 50(4) disclosure duty may well sit with you. The split is set out on the Article 50 transparency page.
- When was it first made available in the Union? Find the evidence now: the launch changelog, the first EU customer contract, the pricing page in the wayback record, the app store release note. This is the fact the whole answer rests on and it gets harder to reconstruct with time.
- Before 2 August 2026? You have until 2 December 2026. On or after? The duty already applies and any unmarked output since placement is exposure.
The four hard cases
Most products answer cleanly. Four situations do not, and they are worth surfacing rather than glossing.
A new version of an old product.A system on the market since 2024 that ships a new model behind the same interface in October 2026 raises the question whether that is a fresh placing on the market or a continuation of the old one. The transitional provision does not answer it. Recital 38 grants the relief to "providers who have already placed their systems on the market before the 2 August 2026" and says nothing at all about new versions, substantial modification or a later change of intended purpose. That silence is the finding, not a gap in this page: there is no continuation rule to rely on, so the reading that treats a materially changed system as newly placed is the one that survives challenge. Mark the new version.
Continuous deployment. A product that ships weekly has no single placement event after the first one if you read the relief broadly, and a new one every week if you read it narrowly. Recital 38 is silent here too, and the relief is drafted as a one-off accommodation for adaptation cost rather than as a rolling entitlement. Teams shipping continuously should assume the narrow reading, because the alternative is betting a middle-tier penalty on an interpretation the text does not support.
Fine-tuned and self-hosted open-weight models. Taking an open-weight model, fine-tuning it and making the result available can make you the provider of a system you did not train from scratch, and the Article 2(12) open-source exclusion does not rescue you here, because it falls away for systems subject to Article 50. If you first made that system available after 2 August 2026, you are outside the transitional period.
Embedded generative features. A generative feature added to a long-standing non-generative product is the case where the continuation reading is least attractive. The generative capability itself was not on the market before 2 August 2026 in any meaningful sense, so treating the feature as covered by relief granted to the host product is optimistic.
What you owe when the date arrives
Article 50(2) requires outputs to be marked in a machine-readable format and detectable as artificially generated or manipulated, using technical solutions that are effective, interoperable, robust and reliable as far as this is technically feasible, taking account of the specificities and limitations of the various content types, implementation costs and the generally acknowledged state of the art.
Those four adjectives are the standard a regulator will measure you against, and they are the reason a visible watermark alone is usually not enough: a caption burned into an image is not machine-readable, and metadata alone is not robust, because it is stripped by ordinary platform re-encoding. The Code of Practice on transparency of AI-generated content, final text 10 June 2026, is voluntary but is the de facto yardstick, because nothing else fills the gap. It expects multi-layered marking where a single technique cannot meet all four criteria, exempts free-form text under 200 tokens and generative systems embedded in closed physical products, and sets a watermark-detection interoperability expectation from 2 February 2027.
What to do with the runway
If you are inside the transitional period, the runway to 2 December 2026 is roughly the lead time for getting marking into a generation pipeline, which makes it adequate rather than generous. Four steps, in the order that removes the most exposure per hour spent.
- Fix the placement date in writing now, with the evidence attached. It is the fact your whole position rests on and it is cheapest to establish today.
- Inventory every output type the system generates. Image, audio, video and text can need different marking techniques, and teams routinely forget that a product they think of as an image tool also generates captions.
- Put the requirement to your vendors in writingif any part of the generation is someone else's. Ask what marking they apply, whether it survives re-encoding, and whether they will commit to the Code of Practice. A vendor that cannot answer in August will not be able to answer in November either.
- Do not let the marking work wait on the high-risk programme. The high-risk regime moved to 2 December 2027 and 2 August 2028, and the full timeline as amended keeps those apart. Marking is the nearer date by more than a year.
If you are outside the transitional period, the position is simpler and less comfortable: the duty applies now, and the first question is not when to build the marking but what your exposure has been since placement.