10 Apr 2026 AI Engineering 6 min read

EU AI Act Article 50 — Transparency by 2 August 2026

The high-risk obligations of the AI Act are being deferred to December 2027. Article 50 is not. From 2 August 2026 providers must mark generated output in machine-readable form and disclose AI interaction; deployers must label deepfakes. We set out what to implement now — marking, interface disclosure, logging — and what the Act does not require.

The Deadline That Did Not Move

Most AI Act news over the past six months has been about postponement. The Digital Omnibus on AI, proposed on 19 November 2025, would push the high-risk obligations for Annex III systems from 2 August 2026 to 2 December 2027. That headline produced a widespread and wrong conclusion: that nothing lands this summer. Article 50 lands. Its duties apply from 2 August 2026, and neither co-legislator proposed moving that date.

Article 50 is not a paperwork obligation. It requires changes in the product: a machine-readable mark on generated output, a disclosure in the interface, a labelling decision in the publishing pipeline. That is engineering work with a fixed delivery date 113 days away. What follows is engineering guidance, not legal advice — for the binding interpretation of a specific case, ask a lawyer.

Untrusted input arrives — treat it as data, not instructions. 1/4

What Article 50 Actually Requires

Article 50 of Regulation (EU) 2024/1689 carries four substantive duties across two roles. Paragraphs 1 and 2 bind the provider — whoever places a system on the market under their own name. Paragraphs 3 and 4 bind the deployer — whoever uses it under their own authority. Paragraph 5 governs the form of the notice. Paragraph 6 confirms these duties are additive to Chapter III and other Union law.

The exemptions are narrow. Paragraph 1 does not apply where the AI nature is obvious to a reasonably well-informed, observant and circumspect person. Paragraph 2 does not apply where a system performs an assistive function for standard editing or does not substantially alter the input data or its semantics. Paragraph 4 reduces the duty for evidently artistic, creative, satirical or fictional works to a disclosure that does not hamper enjoyment.

Article 99(4)(g) caps penalties for breaches of Article 50 at EUR 15 million or 3 percent of total worldwide annual turnover, whichever is higher. Article 99(6) inverts that for SMEs: whichever is lower. Member States set the actual penalties and designate the authority that imposes them.

Provision	Bound party	Trigger	Duty
Art. 50(1)	Provider	System interacts directly with a person	Inform that this is an AI system
Art. 50(2)	Provider	Synthetic audio, image, video or text	Machine-readable mark, detectable output
Art. 50(3)	Deployer	Emotion recognition, biometric categorisation	Inform the persons exposed
Art. 50(4)	Deployer	Deepfake image, audio or video	Disclose artificial generation
Art. 50(4)	Deployer	AI text on matters of public interest	Disclose, unless editorial responsibility
Art. 50(5)	Both	All of the above	Clear, distinguishable, at first exposure, accessible

Marking Is a Provider Duty

Article 50(2) requires that the outputs of generative systems are marked in a machine-readable format and detectable as artificially generated or manipulated. It names no technique. It sets a standard of care instead: effective, interoperable, robust and reliable as far as technically feasible, taking into account the state of the art and the cost of implementation. The duty sits with the provider, including providers of general-purpose models offered as a service.

The second draft of the Code of Practice on Transparency of AI-generated Content, published on 5 March 2026, describes a two-layer approach: cryptographically signed and timestamped provenance metadata, plus a watermark in the signal itself. The layers fail differently. Metadata is precise and verifiable but is stripped by most upload pipelines and by every screenshot. Watermarks survive re-encoding and cropping but degrade under heavy editing and carry far less information.

Content Credentials specification 2.3, published in January 2026 and moving to ISO as ISO/DIS 22144, is the practical container for the metadata layer. Text remains the hard case: token-level watermarks survive light editing and fail under paraphrase and translation. Keep a provider-side generation log — model, version, timestamp, request digest. When marks are stripped, that record is the only thing still answering whether an artefact is yours.

Disclosure Is an Interface Problem

Paragraph 5 is the requirement teams underestimate. The information must be clear and distinguishable, must reach the person at the latest at the time of first interaction or exposure, and must meet applicable accessibility requirements. A dismissible modal at signup satisfies none of that for a widget on a third-party page, a deep link into an existing conversation, or a voice channel where nothing is visible.

We build disclosure as a property of the channel rather than of the session. Every surface that renders model output declares its own notice: a persistent label in chat, a spoken sentence before the first turn in voice, an explicit line in email and ticket replies. The notice reaches assistive technology as text, not as colour or an icon alone. Under paragraph 3 the notice comes before processing starts.

Provider or Deployer Changes Everything

Duties follow the role, and the role is not decided by who wrote the code. A provider places a system on the market under its own name or trademark. A deployer uses it under its own authority. Under Article 25, a deployer that puts its own name on a system, changes its intended purpose, or makes a substantial modification becomes a provider of that system — and inherits paragraph 2 with it.

Most organisations we work with are deployers of a general-purpose model and providers of the application built on it. That combination is the common failure mode. The vendor marks what it generates; the application then composes, crops, converts and re-renders that output, and the mark does not survive the pipeline. Marking must be re-applied where your system emits the artefact. Put that expectation in the upstream contract and verify it in a test.

What the Act Does Not Require

Article 50 does not require a visible watermark on every generated image. It requires a machine-readable mark; the visible label is a separate duty, falls on the deployer, and covers only deepfakes and public-interest text. It does not require disclosure where the AI nature is obvious. It does not reach output that is never published, and it does not restrict what may be generated — that is Article 5 and Chapter III.

It also guarantees nothing. A mark shows that a system claims to have generated a piece of content. It does not show that unmarked content is human, and detection tools do not remove that asymmetry. Article 50 is a disclosure duty on identified operators, not a provenance system for the open web. Compliance says nothing about the GDPR, the DSA or copyright, which apply in parallel under paragraph 6.

What the Digital Omnibus Moved

The Digital Omnibus on AI — COM(2025) 836 final of 19 November 2025 — is still a proposal. The Council agreed its general approach on 13 March 2026; Parliament adopted its negotiating position on 26 March 2026 with 569 votes in favour, 45 against and 23 abstentions. Trilogues are running. None of it is law until they conclude and the final text is published.

What moves is the high-risk regime: Annex III systems to 2 December 2027, product-embedded Annex I systems to 2 August 2028. The Commission tied the deferral to confirmed availability of harmonised standards; both co-legislators prefer fixed dates. For Article 50 the proposal touches one point — a transition for the paragraph 2 marking duty, for systems already placed on the market before 2 August 2026, to 2 February 2027.

National enforcement is unfinished as well. Germany's implementing act had its first Bundestag reading on 20 March 2026 and a committee hearing on 23 March 2026, designating the Bundesnetzagentur as market surveillance authority. The obligations apply whether or not the national authority is operational on day one.

Build the Marking Layer Now

Two things are still outstanding on the regulator's side. The Code of Practice under Article 50(7) is expected to be finalised in early June 2026, and the Commission guidelines under Article 96(1)(d) on the practical implementation of Article 50 have not been published. Both will sharpen the detail. Neither changes the date, and neither is a precondition for the work.

So we build against what is already fixed. One point in the architecture where generated artefacts are marked and logged. A disclosure component owned by the channel. A labelling decision recorded in the publishing pipeline together with the person who made it. Signing keys, rotation and verification treated as production infrastructure.

Our expectation for the second half of 2026 is that marking stops being a compliance artefact and becomes an addressing mechanism. Once every generated artefact carries a signed identity, automated pipelines can decide what to trust before acting on it. The transparency duty is the cheapest version of that infrastructure we will be asked to build.