Article 50 transparency obligations: the AI Act deadline of 2 August 2026 that has not been postponed

The Digital Omnibus political agreement of 7 May 2026 postponed the high-risk AI obligations to 2 December 2027 and 2 August 2028, but it left the Article 50 transparency obligations exactly where they were: they apply from 2 August 2026. If your organisation deploys chatbots, generates synthetic content, or produces deepfakes, that deadline is roughly seven weeks away and it has not moved.

There is a quiet misreading travelling through compliance teams right now. The headlines about the Digital Omnibus all said the same thing: the AI Act is being delayed. Boards heard "delayed" and assumed the whole high-risk machine had been pushed back, which it largely has. The problem is that Article 50 is not a high-risk obligation. It is a separate transparency layer that applies regardless of risk classification, and the Omnibus deliberately left it alone. So while one part of your roadmap just gained sixteen months, the part that touches every customer-facing chatbot and every piece of AI-generated content did not move at all.

This post is about that gap. It is the companion to our practical guide on Article 50 labelling and detection, which walks through the technical marking work itself. Here the focus is narrower and more urgent: what the Omnibus did and did not change, and why 2 August 2026 is still a live date.

What Article 50 actually requires

Article 50 of the AI Act sits outside the risk-based pyramid that most people picture when they think about the regulation. It does not ask whether your system is high-risk, limited-risk, or minimal-risk. It asks a different question: is a person being exposed to AI without knowing it? Where the answer is yes, the Article imposes a transparency duty.

There are four core situations. First, providers of AI systems that interact directly with people, such as chatbots and virtual assistants, must make sure those people are informed they are dealing with a machine, unless it is obvious from the context. Second, providers of generative AI that produces synthetic audio, image, video, or text must mark that output in a machine-readable format so it is detectable as artificially generated. Third, deployers who use AI to create deepfakes must disclose that the content has been artificially generated or manipulated. Fourth, deployers who use AI to generate or manipulate text published to inform the public on matters of public interest must disclose that too.

The reach is wider than most organisations assume. A marketing team using a generative model to draft public-facing copy, a service desk running a customer chatbot, a media outlet producing synthetic voiceovers: all of them sit inside Article 50. The obligation does not wait for a high-risk classification, which is exactly why the Omnibus postponement does not reach it.

The one date that did move within Article 50

There is a single nuance worth getting right, because it is where the genuine relief lives. The general transparency duties under Article 50, including the chatbot disclosure and the deepfake disclosure, apply from 2 August 2026. But the specific machine-readable marking obligation on providers of generative AI under Article 50(2) carries a separate, slightly later timing for content systems that were already on the market.

In practice this means the headline disclosure obligations land on 2 August 2026, while providers of generative systems get until 2 December 2026 to have their machine-readable marking solutions implemented and operating reliably. That is not a postponement granted by the Omnibus. It is the structure that was already written into the timeline. Treating it as a reason to relax is a mistake: four months is a short runway for watermarking, provenance metadata, and detection tooling that has to survive being copied, compressed, and re-uploaded across platforms.

The deadlines side by side

The clearest way to see what the Omnibus did is to put the obligations in one timeline. The high-risk dates moved. Article 50 did not.

The pattern is deliberate. The Omnibus targeted the part of the AI Act that was hardest to operationalise on time, the high-risk classification regime, while leaving the transparency obligations untouched precisely because they are seen as low-burden and citizen-facing. For a fuller view of every milestone, see our overview of the AI Act deadlines for 2026, 2027 and 2028.

Why this catches organisations off guard

The trap is structural rather than careless. Most compliance programmes were built around a single anchor date of 2 August 2026, with high-risk classification as the centre of gravity. When the Omnibus moved that anchor, the natural response was to move the entire programme with it. But Article 50 was never bolted to the high-risk regime. It rode along on the same calendar date by coincidence, not by design, and now that the high-risk date has slipped, Article 50 has been left standing alone on 2 August 2026.

There is a second reason it gets missed. Article 50 obligations are spread across roles. Providers carry the marking duty, deployers carry the disclosure duty, and many organisations are both at once depending on the system. A company can be a deployer of a third-party chatbot and a provider of its own fine-tuned content model in the same week. That role-splitting means responsibility for Article 50 rarely sits with one owner, so it falls between the cracks of a programme organised around high-risk system inventories.

Practical steps before 2 August 2026

The work is narrower than the high-risk programme, which is good news given the timeline. Start by mapping every point where your organisation touches the four Article 50 situations: customer-facing chatbots and assistants, any generative model that produces public content, any deepfake or synthetic media production, and any AI-assisted text published to inform the public. This is a use-case inventory, not a system inventory, and it usually surfaces deployments that the high-risk mapping never caught.

For chatbots, the fix is often the lightest: a clear, early disclosure that the user is interacting with AI, placed where it cannot be missed. For generated content and deepfakes, the disclosure has to be accompanied by the technical marking, which means engaging vendors now on machine-readable provenance, watermarking robustness, and the detection mechanisms providers must make available. Contracts with generative AI suppliers should be checked for whether the marking is delivered by the provider or expected from you.

The supporting Code of Practice on transparency of AI-generated content is still being finalised, with the Commission having run consultation rounds through the first half of 2026. The Code is voluntary, but adherence will be the practical evidence that your marking approach is adequate. Building toward its layered marking expectations now is more efficient than retrofitting later. Do not wait for the final text to start the engineering work, because the obligation itself does not depend on the Code being finished.

Where this fits in the wider picture

It is worth stepping back. The AI Act already has obligations in force today: the AI literacy duty under Article 4 has applied since 2 February 2025, and the prohibited practices since the same date. Article 50 is the next hard edge on the calendar, arriving on 2 August 2026, ahead of the postponed high-risk regime in late 2027. Organisations that read the Omnibus as a general reprieve have the sequence backwards. The obligations that touch the most people, transparency and literacy, are the ones that are live or arriving first. The heavy, system-specific high-risk obligations are the ones that have been pushed out.

If your AI governance roadmap currently has everything clustered around late 2027, the most valuable hour you can spend this month is separating the Article 50 work back out and putting it where it belongs: due in seven weeks, not seventeen months.