Article 50 is in force. The grace period closes 2 December 2026.

Whether the system is placed on the market, put into service, or its output used, in the European Union. A confirmed no leaves Article 50(1) to 50(5) all not engaged. An unsettled answer is not a no: it caps every paragraph at may be engaged, because no verdict can be firmer than the scope question it rests on. Settle that one first. It is usually a contract and traffic question rather than a legal one.

Article 50(1) is a design obligation and it sits on the provider. A system intended to interact directly with natural persons has to be designed and developed so that the people concerned are informed they are dealing with an AI system, unless that is obvious to a reasonably well informed and observant person taking the circumstances and the context of use into account. Two consequences follow from it being a design duty. It is discharged in the product rather than in a policy document, because the Act attaches the obligation to how the system is built. And the carve-out is a judgment about your own interface, which a regulator would read against the least attentive plausible user rather than against the team that shipped it, so the safe position is to write the notice and keep the reasoning. Binding since 2 August 2026, with no transition window.

Article 50(2) requires the provider of a system that generates synthetic audio, image, video or text to mark that output so a machine can detect it as artificially generated or manipulated. The duty is about the output, not about a policy page: a notice a person can read is not a marking a detector can read. It applied from 2 August 2026, and systems already on the EU market before that date carry a transition to 2 December 2026. On what counts as adequate marking there is now something to build against: the Commission published a Code of Practice on Transparency of AI-generated Content on 10 June 2026. It is voluntary and it is not a harmonised standard, so it carries no presumption of conformity, but with no harmonised standard cited for the AI Act it is the most concrete statement of expectations that exists, and adhering to it is treated as one route to demonstrating compliance rather than the only one.

Usually not. Article 50(2) carries an express carve-out where the system performs an assistive function for standard editing, or does not substantially alter the input data provided by the deployer or its semantics. Two cautions before relying on it. Substantially is doing the work in that sentence, and a feature that began as tone correction and has since learned to draft whole paragraphs has probably left the carve-out without anyone deciding that it should. Write down what the system does today, with the date, because that record is the thing you would show.

No, and it is the simpler case. The transition covers systems already on the EU market before 2 August 2026. For a system placed on the market or put into service on or after that date, the Article 50(2) marking duty applied from the day it was placed, so there is no window left to run out. Changelogs, release notes and announcements date this better than memory does, and the date is worth writing down now rather than reconstructing later.

Article 50(1). The systems it catches in practice are customer support assistants, voice agents, intake bots, automated callers and anything else where a person is on the other end of an exchange and may not know what they are talking to. The duty belongs to the provider of the system rather than to every business that switches one on, and it has bound that provider since 2 August 2026: the December date people quote at you is the Article 50(2) marking transition and does not reach this paragraph at all. Where a person would plainly know what they are dealing with, the Act carves the case out, but that is a call about your interface, so record who made it and on what date. A chat window that names itself an AI assistant in its opening line and a voice agent that answers in a human-sounding greeting are not the same case.

Two of them, and they fall on different parties. Article 50(2) is the provider's duty to mark synthetic audio, image, video or text so a machine can detect it as artificially generated. Article 50(4) is the deployer's duty to disclose deepfake image, audio or video, and generated text published to inform the public on matters of public interest. Marketing copy generation, synthetic voice, image editing at scale and generated summaries published under your name are the common cases, and a retrieval system that composes an answer generates text.

Article 50(3), the duty to inform the people exposed to an emotion recognition or biometric categorisation system. It sits with the deployer, not the provider, and it applied from 2 August 2026 with no transition period. It is easy to miss in a product that does this as one feature among many rather than as its headline function, which is where an inventory of AI systems earns its cost. The spelling of categorisation is the Act's own, in the English text of Article 50(3). One boundary worth naming, because this page does not settle it: a disclosure duty assumes the system may be operated in the first place, and where the people exposed are at work or in an education setting, whether it may be run at all is governed by parts of the Act outside Article 50. That is the prior question, and it goes to counsel before any notice gets written.

The deployer. The first subparagraph of Article 50(4) requires the deployer of a system that generates or manipulates image, audio or video content constituting a deep fake to disclose that it has been artificially generated or manipulated. Article 3(60) defines a deep fake as AI-generated or manipulated image, audio or video content resembling existing persons, objects, places, entities or events that would falsely appear to a person to be authentic. If you are only the provider it is your customer's obligation, and worth documenting for them.

No, and this is the case most often reported as an exemption. Where the content forms part of an evidently artistic, creative, satirical or fictional work, the obligation is reduced rather than removed: the existence of generated content still has to be disclosed, in a manner that does not hamper the display or enjoyment of the work. A credit is usually enough. Silence is not.

No. That carve-out belongs to the text limb of Article 50(4) only, the second subparagraph, which covers generated text published to inform the public on matters of public interest and does not apply where the content has undergone human review and a person holds editorial responsibility for the publication. It does not reach deepfake image, audio or video. A broadcaster with full editorial control over a synthetic video still owes the disclosure on that limb.

Article 50(5) governs the manner and the timing of whatever the paragraphs above require. The information has to be given clearly and distinguishably, at the latest at the time of the first interaction or exposure, and it has to conform to the applicable accessibility requirements. It binds the provider and the deployer, whichever of them holds the duty it attaches to. It is the paragraph most often satisfied on paper and failed in the interface.

Article 50 runs to 50(7). Article 50(6) says that 50(1) to 50(4) leave Chapter III and other transparency law untouched, and 50(7) covers codes of practice and was rewritten by the Digital Omnibus on AI. Neither puts a duty on a provider or a deployer, which is why this page lists 50(1) to 50(5) and names that range rather than calling them all five paragraphs of the Article. Read at EUR-Lex on 14 September 2026.

Regulation (EU) 2026/1744, the Digital Omnibus on AI, adopted on 8 July 2026, amending Regulation (EU) 2024/1689. The mechanism is worth knowing if you ever check this against the text yourself: the new schedule is not written into the Omnibus as a free-standing date, it is written by replacing point (c) of Article 113 of the AI Act, the provision that says when Chapter III, the high-risk chapter, starts to apply. So the operative dates are read out of the amended Act rather than out of the amending regulation, which is why a search of the Omnibus alone can leave you thinking the dates moved less than they did. The Article 50 transparency dates were not touched by it. A great deal of guidance was written before July 2026 and still states the superseded schedule as current, so check the publication date on anything you are relying on.

Yes, 2 August 2027, and it is set by Article 111(3) of the AI Act itself rather than by the Digital Omnibus: providers of general-purpose AI models placed on the market before 2 August 2025 have until that date to comply. If you build on someone else's model that obligation is theirs, not yours, and the date is a useful thing to put on the agenda of your next vendor review.

Yes. The Article 50 check at metaminds.store/tools/article-50 runs the same rules this page is built from, over eight questions about one system, in the browser, with no email required. It is engineering triage rather than a compliance determination and it says so: every question carries an explicit I would have to check option, an unsettled scope answer caps everything below it, and two of its five verdicts, not engaged and cannot tell, lead to no paid work at all.

A separate verdict for each paragraph, Article 50(1) through 50(5), rather than one score for the system. The five possible verdicts are engaged, may be engaged, applies in reduced form, not engaged on these answers, and cannot tell from these answers, and each one is printed with the reasoning that produced it and the duty holder it falls on. Three of the five leave something worth working on and two of them do not, which is why a run can end with the tool telling you there is nothing to buy.

No. ISO/IEC 42001 is a management-system standard and the EU AI Act is a regulation. The certificate is evidence of a governed process; it is not a determination that a given system meets Article 50, and no auditor issues one about the Regulation. The two are complementary and neither substitutes for the other. MetaMinds runs ISO/IEC 42001 gap assessment, readiness, control design and evidence preparation; the certificate itself is issued by an accredited certification body after an external audit.

That is a question about machinery rather than about obligations, and the honest answer changes month to month, so MetaMinds keeps a dated register of it at metaminds.store/eu-ai-act-status instead of asserting a state here: harmonised standards, notified bodies and national authorities, each checked at source with the date it was checked. The obligations in Article 50 apply whether or not the machinery around them is finished.

MetaMinds does not publish a fine figure on this page. Penalty exposure under the AI Act depends on the provision engaged and on the member state enforcing it, so a number quoted out of that context is the kind of thing your counsel corrects in front of your board. What is worth stating plainly instead: the AI Office and member-state authorities have held enforcement powers under the Act since 2 August 2026, the same day Article 50 began to apply.