learn-ai-governance-with-phoebe / Practitioner session 3 of 8
Learn AI Governance with Phoebe · Practitioner track · Session 3 of 8

Transparency obligations tell people it is AI

Article 50 is short, cheap to comply with, and easy to forget. It says: when people interact with AI, or see AI-generated content, they should know. Four duties, split between providers and deployers, all triggered at first interaction or exposure. This session works through each one, then Vantage checks whether its candidate chatbot discloses that it is AI.

🟡 Medium AI/ML · risk · compliance The Article 50 duties After b2
0-3 · Welcome 3-18 · Concepts 18-40 · Build-along 40-45 · Q&A
Part 0

The duty that is easy to owe and easy to miss

Transparency is its own obligation, distinct from the risk tier. A limited-risk chatbot owes it. So does a high-risk system - Article 50 rides on top of everything in b2. The duties are simple: disclose that people are dealing with AI, and mark AI-generated content so it can be recognised. They cost little to meet and a lot to ignore, because they are the most visible thing a regulator or a journalist can check.

Read this first Educational, not legal advice. This is a plain-language walk through Article 50, not a substitute for reading Regulation (EU) 2024/1689 and getting advice on your specific product.
Live - presented in session Self-study - full depth after class ★ Try it now Sources covered at the end
★ What you build today The four Article 50 duties, who bears each one (provider vs deployer) and when it triggers, what "machine-readable marking" actually means, the artistic and editorial exemptions - and an audit of one chatbot or GenAI feature against all of it.
Part 1 · the four duties

The Article 50 duties 8 min live

Article 50 sets four transparency duties. Two fall on providers (build it into the system), two fall on deployers (disclose at the point of use). All of them land at the latest at the first interaction or exposure.

50.1 · Chatbot discloses AIProvider builds it in · person told theyare interacting with an AI, unless obviousBearer: PROVIDER 50.2 · Mark generative outputAudio / image / video / text marked asartificially generated · machine-readableBearer: PROVIDER of GenAI 50.3 · Emotion / biometricInform the people exposed to emotionrecognition or biometric categorisationBearer: DEPLOYER 50.4 · Disclose deepfakesDisclose deepfake content is artificiallygenerated / manipulated (lighter for art)Bearer: DEPLOYER 50.5 · Timing: give the disclosure clearly and distinguishably, at the latest at the first interaction or exposure. Two provider duties, two deployer duties - and all of them land no later than first contact.
🔍 Click to zoom - the four Article 50 transparency duties
Live50.1 - the chatbot must disclose it is AI2 min▶
  • The duty (50.1): providers must design AI systems intended to interact directly with people so that those people are informed they are interacting with an AI system - unless it is obvious from the circumstances to a reasonably observant person.
  • Who bears it: the provider builds the disclosure into the system.
  • When: at the latest at the first interaction (50.5).
  • In practice: a one-line notice at the top of a chat - "You are chatting with an AI assistant" - usually satisfies it.
★ Try it now (any chat AI)Explain the four transparency duties under Article 50 of the EU AI Act, and for each tell me whether the provider or the deployer bears it and when it must be given. Then tell me what a candidate-facing recruitment chatbot must disclose.
Live50.2 - mark AI-generated content, machine-readably2 min▶
  • The duty (50.2): providers of generative AI must ensure outputs - audio, image, video, or text - are marked as artificially generated or manipulated in a machine-readable format.
  • The standard: the marking must be effective, interoperable, robust, and reliable, as far as is technically feasible.
  • Who bears it: the provider of the generative system.
  • Why machine-readable: so downstream platforms and tools can detect the marking automatically, not just a human squinting at a caption.
Live50.3 and 50.4 - the deployer's two disclosures3 min▶
  • Emotion recognition / biometric categorisation (50.3): deployers of these systems must inform the people exposed to them (and process personal data lawfully). The duty is on the deployer, because they are the ones running it on real people.
  • Deepfakes (50.4): deployers who create or use deepfake image, audio, or video content must disclose that it is artificially generated or manipulated. A lighter regime applies to clearly artistic, creative, or satirical work - the disclosure just cannot spoil the enjoyment. For AI-generated text published to inform the public on matters of public interest, disclosure is required unless the content went through human review with editorial responsibility.
  • Timing (50.5): all disclosures given clearly and distinguishably, at the latest at the first interaction or exposure.
Part 2 · in practice

Doing it in practice 8 min live

The duties are clear on paper. In a real product they turn into two concrete things on every piece of AI content: a machine-readable mark buried in the file, and a visible disclosure a human can actually read.

GenAI modelproduces output Machine-readablewatermark / provenance Visible disclosure"AI-generated" label Person seesboth signals 50.2 · for machines 50.1 / 50.4 · for humans Two marks on one artefact: a machine can detect it, and a human can read it. You usually need both.
🔍 Click to zoom - AI content carries a machine-readable label and a visible disclosure
LiveWhat "machine-readable marking" means3 min▶
  • The idea: a signal embedded in the content itself that software can detect - a watermark in an image or audio track, or provenance metadata (for example content-credential data) attached to the file.
  • The bar (50.2): effective, interoperable, robust, and reliable, as far as technically feasible. Interoperable so other platforms can read it; robust so it survives ordinary editing.
  • Not the same as a caption: a visible "AI-generated" label helps humans (50.1 / 50.4), but 50.2 specifically wants a machine to be able to tell. In practice you carry both.
LiveThe artistic and editorial exemptions2 min▶
  • Artistic / satirical (50.4): where a deepfake is part of an evidently artistic, creative, satirical, or fictional work, the disclosure obligation is lighter - you disclose in a way that does not hamper the display or enjoyment of the work.
  • Editorial responsibility (50.4): AI-generated text published to inform the public on matters of public interest must be disclosed - unless it underwent human review and a person or organisation holds editorial responsibility for it.
  • The trap: "it is creative" is not a blanket exemption. The lighter regime is narrow; a recruitment chatbot or a marketing deepfake does not get to hide behind it.
LiveLimited-risk can ALSO be high-risk3 min▶
  • Transparency is a distinct layer: Article 50 is not a risk tier - it is an obligation that can attach to a system regardless of its tier. A system can be high-risk AND carry Article 50 duties at the same time.
  • Why it trips people: teams think "we are just limited-risk, so all we owe is a chatbot notice." But a high-risk system that happens to chat with candidates owes the full b2 programme and the Article 50 disclosure.
  • The honest test: classify for the tier (b1), build the tier's obligations (b2), then ask separately: does anyone interact with this AI, or see content it generated? If yes, Article 50 is live too.
Real world

Vantage's candidate chatbot must disclose it is AI. Vantage, the Singapore firm shipping its high-risk CV screener into the EU, adds a chatbot that answers candidate questions during the application. Because a person interacts with it directly, Article 50.1 applies: it must open with a clear line that the candidate is talking to an AI assistant. The system is high-risk (b1) AND owes this transparency duty - the two stack. If the tool also generated any synthetic content shown to candidates, 50.2 marking would apply on top.

Cheap insurance Transparency is the lowest-cost, highest-visibility obligation you owe. A single clear disclosure line and a provenance mark cost almost nothing and remove the easiest complaint anyone can make about your system. Do it even where you are unsure it is strictly required.
Build-along · take it further

Audit a chatbot / GenAI feature against Art 50 ★ 10 min

Take one real AI feature that talks to people or generates content, and run it against the four duties.

Is AI disclosed? If people interact with the system, check whether a clear "this is an AI" notice appears at first interaction (50.1). If not, it is obvious enough to a reasonable person - or is it a gap?

Is content labelled? If the feature generates audio, image, video, or text, check for a machine-readable mark (50.2) and a visible disclosure. Note which is missing.

Emotion / deepfake? If you run emotion recognition or biometric categorisation (50.3), or create deepfakes (50.4), confirm the people exposed are informed and the content is disclosed.

Check the timing + the stack. Confirm each disclosure lands at first interaction (50.5). Then ask: is this feature also high-risk? If so, Article 50 stacks on top of the b2 programme.

★ Vantage after b3 Vantage's candidate chatbot now opens with a plain "you are chatting with an AI assistant" line, satisfying Article 50.1 at first interaction. The team has confirmed the disclosure stacks on top of the high-risk programme rather than replacing it, and logged that any synthetic content shown to candidates would need 50.2 marking. Next, b4 turns to general-purpose AI models.
Homework

Before session b4 ◐ 40 min total

Source material

Official sources covered

Taught from the EU AI Act text. This page covers ~80% of the working content on transparency.

✓EU AI Act Article 50 (50.1 - 50.5)Part 1 · the four duties, who bears each, and the timing rule
✓Article 50.2 (machine-readable marking of generative output)Part 2 · watermark / provenance and the effective-robust-reliable bar
✓Article 50.4 (deepfake + editorial exemptions)Part 2 · the artistic / satirical and human-review carve-outs
◐Transparency vs risk tier (a system can be high-risk AND owe Art 50)Part 2 · why the duty stacks on top of the b2 programme
Check yourself

Three questions before you go 🎯 ◐ 90 seconds

1 · Under Article 50.2, generative AI outputs must be marked as artificially generated in what form?

Article 50.2 requires providers of generative AI to mark outputs (audio, image, video, text) as artificially generated in a machine-readable format that is effective, interoperable, robust, and reliable so far as technically feasible.

2 · When must an Article 50 disclosure be given at the latest?

Article 50.5 says the disclosures must be provided clearly and distinguishably, at the latest at the time of the first interaction or exposure.

3 · A high-risk hiring system adds a candidate chatbot. What does it owe on transparency?

Transparency is a distinct obligation. A system can be high-risk AND carry Article 50 duties - the chatbot must disclose it is AI (50.1), and that stacks on top of the full high-risk programme, it does not replace it.

Practitioner session 3 cheat sheet · pin this

Four duties (Art 50)Chatbot discloses AI, generative output marked, emotion/biometric subjects informed, deepfakes disclosed.
50.1 - chatbotProvider builds in a notice that a person is interacting with AI, unless obvious.
50.2 - mark outputProvider marks generative output machine-readably: effective, interoperable, robust, reliable.
50.3 / 50.4 - deployerDeployer informs people exposed to emotion/biometric systems, and discloses deepfakes.
Timing (50.5)Clearly and distinguishably, at the latest at first interaction or exposure.
Machine-readable = detectableWatermark or provenance metadata a machine can read - not just a human caption.
Exemptions are narrowArtistic/satirical is lighter; editorial-reviewed text is carved out. Not a blanket escape.
It stacksA system can be high-risk AND owe Article 50. Transparency does not replace the programme.