Skip to content
MetaMinds
Tools

Which parts of Article 50 actually reach your system?

Article 50 has five paragraphs. They put different duties on providers and on deployers, and only one of them carries a transition rather than applying outright. Eight questions is usually enough to tell them apart.

Scale
Eight questions
Cost
Free
Sign-up
None
Stored
Nothing
Transition window

33days until 2 December 2026

2 Aug 20262 December 2026

Counted on 2026-10-30, Europe/Brussels. Systems generating synthetic content, already on the EU market before 2 August 2026, have to carry machine-readable marking.

The part most guidance still gets wrong

Four of the five paragraphs in Article 50 had no transition at all. They applied from 2 August 2026. Only the machine-readable marking duty in 50(2), and only for systems already on the market before that date, runs to December.

A great deal of published material describes the December date as though it covered the whole of Article 50, or as though it were the high-risk regime. It is neither. The questions below separate them.

Read this before the questions

This is an engineering triage against the text of Article 50: which paragraphs engage for a system like yours, who holds the duty, and which date applies. It runs on what you type, so it cannot see your system, your contracts or your market, and the compliance determination is your counsel's to make. Take the result to them: it is a real URL, so you can send it without MetaMinds in the middle of that conversation. MetaMinds is not a law firm and does not provide legal services in any jurisdiction, and nothing here is a substitute for the advice or services of an attorney.

  1. 01Which are you, for this system?

    Article 50 puts different duties on the provider and the deployer, and the same organization is often both for different systems. The provider develops it or puts it on the market under its own name or trademark. The deployer uses it under its own authority.

  2. 02Is the system placed on the market, put into service, or is its output used, in the European Union?

    The Act reaches non-EU organizations. If people in the EU use the system, or the output produced by it is used in the EU, it is generally in scope regardless of where you are established.

  3. 03Does the system interact directly with people?

    A chat assistant, a voice agent, an automated caller. The question is whether a person is on the other side of an exchange with it.

  4. 04Does it generate or manipulate audio, image, video or text output?

    Synthetic content of any kind, including text. A retrieval system that composes an answer generates text. Note the third option: Article 50(2) does not reach systems that only assist with standard editing or do not substantially alter what the user gave them.

  5. 05When was this system first placed on the EU market or put into service?

    This is the question that decides whether the December 2026 transition applies to you, and it is the one most often answered from memory. Release notes, changelogs and press announcements settle it better than recollection.

  6. 06Does it perform emotion recognition or biometric categorisation?

    Inferring emotional state, or sorting people into categories from biometric data. This is a deployer duty and it is easy to miss in a product that does it as one feature among many.

  7. 07Does the system generate or manipulate image, audio or video that resembles real people, places or events, and would falsely appear authentic?

    This question is about image, audio and video only, because that is the limb of Article 50(4) it maps to. Text is handled by the next question. Examples: synthetic voice of a named individual, a generated image of a real place, a video of someone who did not say the thing.

  8. 08Is generated text published to inform the public on matters of public interest?

    News, current affairs, public health, elections. If it is, whether a human holds editorial responsibility for the published output changes the answer, which is why both parts are asked together.

0 of 8 answered

What this tool does not do

  • It does not tell you whether you are compliant. It tells you which paragraphs are worth reading closely.
  • It does not cover the rest of the Act. Prohibited practices under Article 5, the high-risk regime and the GPAI obligations are separate questions with separate dates.
  • It assumes you are not law enforcement. Article 50 carries carve-outs, across several of its paragraphs, for AI systems authorised by law to detect, prevent, investigate or prosecute criminal offences, subject to safeguards for the rights of third parties. If that describes your deployment, the answers above are not written for you and the carve-outs are worth reading in the text itself.
  • It does not store your answers, set a cookie, or ask for an email address. The result lives in your own URL and nowhere else.
  • It does not replace counsel, and the two places it is most likely to be too coarse are the obviousness carve-out in 50(1) and what counts as editorial responsibility in 50(4).

The full timeline, including which deadlines the July 2026 Digital Omnibus moved and which it did not, is on our Article 50 page.

Questions

Article 50, answered paragraph by paragraph.

The date, the duty holder and the carve-out for each paragraph, written so one answer quoted on its own is still correct. These are the same strings this page puts in its structured data.

When did EU AI Act Article 50 start to apply?

Article 50 transparency obligations have been in force since 2 August 2026, and four of its five paragraphs had no transitional period at all: they applied from that date. The single exception is the machine-readable marking duty in Article 50(2), and only for providers of synthetic-content systems that were already on the EU market before 2 August 2026, whose transition runs to 2 December 2026. A system placed on the market on or after 2 August 2026 had the marking duty from the day it was placed there.

What is the 2 December 2026 Article 50(2) deadline, exactly?

2 December 2026 is the end of a four-month transitional period for the machine-readable marking of synthetic content, and it reaches one population only: providers of AI systems that generate synthetic audio, image, video or text and that were already placed on the EU market before 2 August 2026. It is not a deadline for the whole of Article 50, it is not the high-risk regime, and it does not give a new system launched after 2 August 2026 any extra time. A great deal of published guidance describes it as though it were one of those three.

Who holds each Article 50 duty, the provider or the deployer?

It changes by paragraph, which is the single most common way a reader gets this wrong. Article 50(1), informing a person that they are interacting with an AI system, is a provider duty. Article 50(2), marking synthetic output as machine-detectably artificial, is a provider duty. Article 50(3), informing people exposed to emotion recognition or biometric categorisation, is a deployer duty. Article 50(4), disclosing deepfakes and certain published text on matters of public interest, is a deployer duty. A firm that is the deployer of one system and the provider of another holds different obligations on each.

Does Article 50(2) apply to a chatbot or a RAG system that only generates text?

Article 50(2) covers synthetic content of any kind, and text is included, so a retrieval system that composes an answer generates text within its scope. The paragraph carries an express carve-out: the marking obligation does not apply 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. Grammar correction, reformatting and tone adjustment sit inside that carve-out; the word doing the work is "substantially", and a feature that began as tone correction and has since learned to draft whole paragraphs has probably left it without anyone deciding that it should.

Is my AI system high-risk under the EU AI Act, and does that apply now?

Not yet, on the dates. Annex III stand-alone high-risk obligations apply from 2 December 2027 and Annex I high-risk obligations embedded in regulated products apply from 2 August 2028, both deferred by Regulation (EU) 2026/1744, the Digital Omnibus on AI, which was published in the Official Journal on 24 July 2026 and has been in force since 27 July 2026. The live near-term pressure is transparency and disclosure under Article 50, not the full high-risk conformity regime, and material implying otherwise is usually written against the pre-Omnibus schedule.

Can somebody do an EU AI Act conformity assessment for a high-risk system today?

Not through a notified body, because as of 5 September 2026 there were zero bodies notified under Regulation (EU) 2024/1689 in NANDO, the Commission's own register, filtered to that legislation with notification status Active. No harmonised standards had been cited in the Official Journal for the AI Act either. That is a real constraint on what conformity work can be completed today, and MetaMinds maintains a register of what does and does not exist yet rather than asserting readiness the machinery cannot deliver.

What does an AI system have to disclose under Article 50(1)?

That a person is dealing with an AI system rather than a human, built into the system by its provider, for systems intended to interact directly with natural persons. The paragraph carves out cases where that is obvious to a reasonably well informed and observant person, and whether a given interface clears that bar is a judgment about the interface rather than a fact a form can settle. A regulator would test it against the least attentive plausible user, not the most attentive.

Are there exemptions to Article 50 for law enforcement?

Yes. Article 50 carries carve-outs across several of its paragraphs for AI systems authorised by law to detect, prevent, investigate or prosecute criminal offences, subject to safeguards for the rights of third parties. This tool assumes the reader is not law enforcement and its answers are not written for that case; where it describes your deployment, the carve-outs are worth reading in the text of the Regulation itself.

Does this Article 50 checker store my answers or ask for an email?

No. The tool is a plain HTML form submitted by GET, so the answers become query parameters in your own address bar and the page renders the result from them. There is no database write, no analytics event and no hidden identifier, which is a claim MetaMinds can make because there is no code that could do it. The result is a real URL, so it can be shared with counsel without MetaMinds being in the middle of that conversation, and the whole tool works with JavaScript disabled.

What is this for, and what should I do with the result?

It is an engineering triage: a short list of which paragraphs of Article 50 are worth taking to counsel and which are not, with the duty holder and the applicable date on each. Take it to them, which is what the shareable URL is for. The two places it is most likely to be too coarse are the obviousness carve-out in 50(1) and what counts as editorial responsibility in 50(4). MetaMinds is not a law firm and does not provide legal services in any jurisdiction, and nothing here is a substitute for the advice or services of an attorney.

What happens after the triage, if something is engaged?

An EU AI Act readiness assessment at MetaMinds is $4,900 over one week and produces the evidence behind each engaged obligation: what the system does, what is disclosed and where, what marking exists in the output, and what a competent authority would be shown if it asked. Where the triage finds nothing engaged, this page says so and offers nothing, which is why three of its five verdicts lead nowhere.

Two pages go deeper: what Article 50 requires and what the evidence looks like, and a maintained register of which parts of the Act's machinery exist yet, checked at primary source with dates.