In one paragraph

Most of the coverage of this summer's Digital Omnibus focused on what was postponed. The high-risk classification regime moved from August 2026 to December 2027, and to August 2028 for AI embedded in regulated products. That deferral is real, and for anyone racing towards an Annex III conformity assessment it is a genuine reprieve. It also created a false sense of general relief, because Article 50 was not deferred at all. The transparency and disclosure regime took effect on 2 August 2026 and applies today. It reaches far more organisations than the high-risk regime ever would, since it is triggered by whether a person can tell they are dealing with an AI system, and not by what that system decides. Run a customer-facing chatbot, a synthetic voice agent, an emotion-detection tool, or any pipeline that publishes generated text or media, and you are inside a live obligation. The board question has changed. It is no longer "are we high-risk". It is "can we evidence that every person interacting with our AI knew they were".

What actually changed on 2 August 2026

The AI Act's obligations were always going to arrive in tranches. Prohibited practices under Article 5 bit first, in February 2025. General-purpose AI model obligations followed in August 2025. The August 2026 tranche was supposed to carry the high-risk regime and the transparency regime together.

The AI Omnibus split them apart. Chapter III enforcement for stand-alone high-risk systems under Annex III moved to 2 December 2027, and high-risk AI embedded as a safety component in products already covered by Union harmonisation legislation under Annex I moved to 2 August 2028.

Article 50 stayed where it was.

That distinction matters more than it first appears, because the two regimes catch completely different populations. The high-risk regime is narrow and definitional. Your system is in scope because it performs a listed function in a listed domain, such as biometric identification, critical infrastructure, employment decisions or credit scoring. Most organisations, on inspection, found they had one or two such systems, or none at all.

Article 50 is broad and behavioural. It applies because of how a person encounters your system. A customer service chatbot that answers questions about delivery times is not high-risk under any reading of Annex III, and it sits squarely inside Article 50(1).

This is why the undisclosed AI interaction risk is now a live exposure for organisations that had otherwise concluded the AI Act was somebody else's problem.

The four obligations, and who carries each

Article 50 is four duties, not one, and they fall on different parties. That is where most implementation programmes go wrong, because the organisation running the chatbot is frequently not the organisation that built it.

Under Article 50(1), systems intended to interact directly with natural persons must be designed and developed so those persons are informed they are interacting with an AI system. This falls on the provider. It is a design duty, attaching to whoever places the system on the market, and not to whoever switches it on.

Under Article 50(2), providers of generative systems must mark outputs in a machine-readable format and ensure they are detectable as artificially generated or manipulated. Also the provider. This is a technical watermarking and provenance obligation, invisible to the user.

Under Article 50(3), if you operate a system that infers emotion or categorises people biometrically, you must inform the people exposed to it. This one falls on the deployer, and it does so even when the system was bought in.

Under Article 50(4), deployers publishing AI-generated or manipulated image, audio or video content constituting a deepfake must disclose it. The same applies to AI-generated text published to inform the public on matters of public interest, unless the content underwent human review and a natural or legal person holds editorial responsibility for it.

The provider and deployer split is the commonest structural failure in Article 50 programmes. An organisation buys a conversational AI platform, configures it, brands it, and deploys it on its own website. It then assumes the vendor carries the transparency duty because the vendor built the model. Under 50(1) that may well be right. Under 50(3) and 50(4) the duty belongs to the deployer and cannot be contracted away, and if the organisation has modified the system substantially or put its own name on it, it may have become a provider in its own right for AI Act purposes. Answer that question before a market surveillance authority asks it. It is as much a vendor governance question as a compliance one.

The exemptions, and the one most organisations will misread

Each obligation carries carve-outs. They are narrower than they look.

Article 50(1) does not apply where disclosure would be obvious to a reasonably well-informed, observant and circumspect person, taking into account the circumstances and the context of use.

That is the exemption that will be over-claimed, and it is worth being blunt about why. "Obvious" is assessed against a legal standard of a reasonably informed observer in context. It is not measured against what your product team believes is obvious, and not against what your most sophisticated users would infer. A chat widget in the corner of a website is not self-evidently an AI system to someone who has never used one, and plenty of organisations staff those widgets with humans. A synthetic voice on an outbound call is close to the clearest case against obviousness, since the entire commercial point of high-quality voice synthesis is that it does not announce itself. If your compliance position rests on obviousness, write down the reasoning, name the assumed user, and have somebody outside the product team test it. An undocumented judgement is not a defence. It is a hope.

Article 50(2) does not reach systems performing an assistive editing function, or systems that do not substantially alter the input data provided by the deployer. Grammar correction is the canonical example. The line is "substantial alteration", and generative rewriting sits on the wrong side of it.

Article 50(4) treats creative work differently. Where deepfake content forms part of an evidently artistic, creative, satirical, fictional or analogous work, the transparency obligation is limited to disclosing the existence of generated content in a manner that does not hamper the display or enjoyment of the work. That is a proportionality carve-out and not an exemption.

Article 50(4) also carves out AI-generated text that has undergone a process of human review or editorial control, where a natural or legal person holds editorial responsibility for the publication. Note both limbs. Review and responsibility. A workflow where a human clicks approve without meaningful scrutiny satisfies neither.

Law enforcement carve-outs run through all four. They are narrow and purpose-bound.

One deadline is still ahead of you

The Omnibus did make one concession inside Article 50. Providers of generative AI systems already placed on the market before 2 August 2026 received a grace period on the Article 50(2) machine-readable marking duty, running to 2 December 2026. Systems placed on the market on or after 2 August 2026 get nothing.

If your organisation provides a generative system that predates August, that is the live deadline, and it is roughly fourteen weeks out at the time of writing. If your organisation only deploys such systems, it becomes a supplier assurance question instead. Your provider's watermarking readiness is now something to ask about in writing, because the marking is what makes your own downstream disclosures verifiable.

Two things exist to help, and one changes your evidential position

The Commission published guidelines on the transparency obligations for providers and deployers, most recently updated on 6 August 2026. They are the authoritative reading of scope and the natural starting point for a gap assessment.

The Code of Practice on Transparency of AI-generated Content was published separately on 10 June 2026, and had attracted roughly 190 signatories by late July. Adherence is voluntary. The evidential consequence is not symmetric, though: signatories can point to the Code's measures as a recognised, EU-wide practical framework for demonstrating compliance with Article 50, while organisations taking an alternative route must demonstrate the adequacy of their own approach to market surveillance authorities on their own account.

That is a governance decision with a real cost attached to either answer, and it belongs at the board and not in the engineering backlog. Signing commits you to a defined set of measures. Declining commits you to building and maintaining your own compliance argument. Neither is wrong. Drifting into the second by default is.

What it costs to get wrong

Breaches of Article 50 sit in the AI Act's general non-compliance tier under Article 99(4), at up to €15 million or 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. Article 50 is named there explicitly, at point (g).

For scale, prohibited practices under Article 5 attract up to €35 million or 7%. Supplying incorrect, incomplete or misleading information to authorities attracts up to €7.5 million or 1%. For SMEs and start-ups each cap applies at the lower of the amount and the percentage, which is one of the few places the Act scales to firm size.

The 3% tier never made the headlines that the 7% tier did, which is precisely why it under-registers at board level. For most organisations of any size it is still materially larger than the cost of doing the disclosure work properly.

What the board should be asking for

Five questions, each answerable with an artefact instead of a reassurance.

  • Where is the inventory of systems that interact with people, publish generated content, or infer emotion? Not the full AI inventory. The Article 50-relevant subset, with each system mapped to whichever of the four paragraphs catches it.
  • For each system, are we the provider, the deployer, or both? Written down, with the reasoning, and reviewed by somebody who understands that substantial modification or own-branding can convert a deployer into a provider.
  • Where are we relying on the "obvious" exemption, and who tested that assumption?
  • What is our position on the Code of Practice, and who decided it? A decision with a date and an owner, not a default.
  • How would we evidence, twelve months from now, that a specific user on a specific date was told they were interacting with AI?

That last question separates a disclosure that exists from a disclosure that can be proven. It is usually the one nobody has an answer to.

Evidencing it, not just doing it

Article 50 compliance fails audit in a characteristic way. The disclosure is present in the product today, and there is no way to demonstrate it was present eighteen months ago when the complaint arose.

Article 50(5) requires the information to be provided to the natural persons concerned in a clear and distinguishable manner, at the latest at the time of the first interaction or exposure. "At the latest at first interaction" is an evidential trigger as well as a design constraint. It is a point in time you may later need to reconstruct.

Three practices close that gap, and none of them is expensive.

Version the disclosure text as a controlled artefact, with change history, the way a privacy notice is versioned. It is a compliance statement and not UI copy, and it should not be editable in a CMS without a trail.

Log the disclosure event, and not merely the interaction. A session record capturing which disclosure version was served, and when, is what turns an assertion into evidence.

Put the provider and deployer determination, and the "obvious" reasoning, in the risk register, dated and owned, so the judgement stays reviewable instead of tacit.

This is the same evidential discipline that ISO/IEC 42001's Statement of Applicability demands of control decisions generally. The decision is not the artefact. The record of the decision is.

Where this sits against your existing register

For organisations already running an AI governance programme, Article 50 is not a new workstream. It is a set of controls attaching to entries most registers already carry: undisclosed interaction, synthetic content provenance, and the transparency obligations sitting alongside them. The work is mapping and not building.

For organisations that concluded in July that the Omnibus had bought them until December 2027, it is a live obligation that has been in force for three weeks, and the gap assessment starts with an inventory.

The high-risk regime was deferred. Transparency was not.