← Back

Guide · Published November 14, 2026 · Reviewed November 14, 2026 · Ingenious AI team · Not legal advice

← All guides

EU AI Act Article 50 Transparency Duties: What Must Be Disclosed, and Since When

The short answer: The EU AI Act's Article 50 transparency obligations have applied since August 2, 2026. They are interaction and content duties, not decision duties: providers must make sure people are informed they are interacting with an AI system (unless that is obvious), mark AI-generated or manipulated content in a machine-readable format, and deployers must inform people exposed to emotion-recognition or biometric-categorisation systems and disclose deepfakes and certain AI-generated public-interest text. The kit's checker gives the same Article 50 result whatever the AI's role in a decision, because these duties do not depend on it.

Check your duties and dates — free Build your first notice — free

What is Article 50? It is the transparency article of Regulation (EU) 2024/1689, the EU AI Act, as carried in the kit's checker data with its dates: Article 50 obligations apply from August 2, 2026, and systems placed on the market before that date have until December 2, 2026 for the machine-readable marking duty. The separate high-risk deployer duties for decision systems were postponed to December 2, 2027 — see EU AI Act High-Risk Decision Systems.

Duty one: tell people they are interacting with AI

Providers of AI systems intended to interact directly with people must make sure the person is informed they are interacting with an AI system, unless that is obvious (Article 50(1)). The 'unless obvious' carve-out is narrower than it sounds in a decision context: a screening portal, an interview tool, or a support chat that decides something about the person is not obviously an AI interaction merely because it feels automated. When in doubt, the disclosure is a sentence — and a dated record that it was shown.

This is the closest EU analogue to the pre-use notices the kit generates for the US states, and it shares their logic: the information arrives at the point of interaction, before the system has done its work on the person. The kit's pre-use notice shape — business, tool, kind of decision, correction and review contacts — exceeds a bare Article 50 interaction notice, which is fine: nothing in the checker data suggests over-disclosure is the risk to manage.

Duty two: mark AI-generated content so machines can read it

Providers must mark AI-generated or manipulated content in a machine-readable format (Article 50(2)). The marking is for systems, not just humans: downstream tools should be able to detect that content was AI-produced. Systems placed on the market before August 2, 2026 have until December 2, 2026 for this duty — a short, specific grace period for this one obligation, not for Article 50 as a whole.

For a business whose 'content' includes decision documents — explanations, notices, screening summaries — the practical question is which outputs count as AI-generated or manipulated content in your stack, and whether the marking survives your export path. A mark applied in the generating tool but stripped by your PDF export is a workflow fact worth discovering in testing, not in an audit. The content-marking duties get their own walkthrough in AI Content Marking and Deepfake Disclosure Duties.

Worked example: Alpenrose Learning Group

Example — fictional business, real duty structure. Alpenrose is a fictional education provider running an AI-assisted admissions and learner-support operation that touches EU applicants. One month after Article 50 applies, its transparency record shows:

Duties three and four: emotion recognition, deepfakes, public-interest text

Deployers of emotion-recognition or biometric-categorisation systems must inform the people exposed to them (Article 50(3)). This is a deployer duty: the business running the interview-analysis or sentiment tool owes the disclosure, even though it did not build the system. Hiring stacks are the obvious exposure point — any tool that infers emotion or categorises biometrics in an interview or assessment triggers the informing duty for the people put through it.

Deployers must also disclose deepfake content, and AI-generated text published to inform the public on matters of public interest (Article 50(4)). Neither duty is a decision duty, but both land on the same communications teams who own notices. The common thread across Article 50 is that the person encountering the AI — interacting with it, being categorised by it, or reading its output — is told what they are encountering.

Why the checker gives one Article 50 answer for every decision role

The kit's free checker asks three questions — jurisdiction, decision type, and how big the AI's role is — and for Colorado, California, and Connecticut the third answer changes the result. For the EU it does not, and the checker says why: the Article 50 transparency duties do not depend on how big the AI's role in a decision is; they attach to interaction and content. A business cannot argue its way out of Article 50 by showing a human made the final call, because Article 50 was never about the final call.

What does depend on the decision area is the next clock. Employment and worker management, education access, and creditworthiness evaluation sit on the Annex III high-risk list, and their deployer duties now apply from December 2, 2027. Article 50 is the duty you already owe; Annex III is the duty you are building toward. Keeping both dates in one tracker — as the kit's record does — stops the 2027 US dates from crowding out the 2027 EU one.

A note on penalties and on the GDPR, stated carefully

The kit's checker data records the enforcement context for Article 50: breaches can be fined up to 15 million euros or 3% of worldwide turnover (Article 99). This guide repeats that figure only because the kit's own verified checker data states it; it is context for prioritisation, not a prediction about any business's exposure, and it is the only penalty figure this guide uses.

The checker also carries a pointer this guide preserves: the GDPR (Regulation (EU) 2016/679, Article 22) gives people rights over decisions based solely on automated processing. It is a different regime and outside the checker — and outside the kit, whose record covers the Colorado, California, and Connecticut duties. A business with EU decision-making needs both conversations; this tool supports the notice-and-record half of the US one and tracks the EU dates alongside. The relationship between the two regimes is explored in GDPR Article 22 vs ADMT Notices.

  1. List every point where a person interacts directly with an AI system, and put the Article 50(1) disclosure at each one
  2. Confirm AI-generated or manipulated content is marked machine-readably — and test that the mark survives your exports
  3. Audit your stack for emotion-recognition or biometric categorisation; if present, inform the people exposed
  4. Add deepfake and public-interest AI text disclosures to the publishing workflow, not as an afterthought
  5. Record disclosures with dates in the same tracker as your US duties — and diarise December 2, 2027 if you decide in an Annex III area

Sources, as carried in the kit's own data: the duty summaries in app/rules.py and the checker data in app/checker_data.py, verified against primary or authoritative sources on October 28, 2026. Run the free duties checker for your jurisdiction, decision type, and the tool's role, and compare the kit's plans for the record, builder, and review queue. More from Ingenious AI is on the hub.

Build your first notice — free

Related guides

See all ADMT guides for the full set.

Frequently asked questions

When did the EU AI Act Article 50 duties start?

August 2, 2026. Systems placed on the market before that date have until December 2, 2026 for the machine-readable content-marking duty specifically.

What are the four Article 50 transparency duties?

Inform people they are interacting with an AI system (unless obvious); mark AI-generated or manipulated content in a machine-readable format; inform people exposed to emotion-recognition or biometric-categorisation systems; and disclose deepfakes and AI-generated text published to inform the public on matters of public interest.

Does Article 50 depend on the AI deciding the outcome?

No. The kit's checker notes that Article 50 duties attach to interaction and content, not to the size of the AI's role in a decision — so the checker returns the same Article 50 result for every role answer.

What fine does the kit's data record for Article 50 breaches?

Up to 15 million euros or 3% of worldwide turnover, under Article 99, as stated in the kit's checker data.

Is the GDPR part of the ADMT Notice Kit?

No. The checker flags GDPR Article 22 as a separate pointer — a different regime, outside the checker and the kit, which covers Colorado, California, and Connecticut notice and record duties.

Build your first notice — free