Skip to main content
Praxikon
Free check
Article 50 of 11344%
Nederlands

Article 50: Transparency obligations for providers and deployers of certain AI systems

Article 50 of the AI Act has required transparency about AI since 2 August 2026. Providers make sure a chatbot discloses that it is AI and mark generated content in a machine-readable way; deployers inform people about emotion recognition and biometric categorisation and disclose deep fakes and AI text on matters of public interest. Only the marking in paragraph 2 has a transition until 2 December 2026, for systems already on the market before 2 August 2026.

, AI and privacy legal adviserobligations checked against the source on
Check your own case for free

Praxikon tracks Article 50 (Transparency obligations for providers and deployers of certain AI systems) under the EU AI Act, citing the source for every statement.

EU Official:
ActiveApplies since 2 Aug 2026
Title IV: Transparency ObligationsApplies from: 2 August 2026

Official text

||

Source: EUR-Lex, Regulation (EU) 2024/1689. Text reproduced verbatim.

Download AI Act (PDF) →

Decision tree

Which transparency obligation applies to your application?

Does your application do more than one thing, for example talk and create images? Then walk through the tree per function.

Question 1 · Art. 50(1) to (4)

What does the application do?

Choose the function you want to check now. Obligations can apply at the same time (Guidelines para. 8). Purely personal use and pure research fall outside Article 50 (Guidelines para. 19, 21).

Full diagram
  1. 1

    What does the application do?

    Art. 50(1) to (4)

    • Talks or communicates with people → Question 2
    • Creates text, images, audio or video → Question 6
    • Recognises emotions or categorises people biometrically → Question 16
  2. 2

    Can people communicate directly with the system, with the system itself responding or acting?

    Art. 50(1) · Guidelines para. 30, 31

    • Yes → Question 3
    • No → Outcome D
  3. 3

    Are you the provider: do you develop the system or have it developed, and place it on the market or put it into service under your own name or trademark?

    Art. 3(3) · Art. 50(1) · Guidelines para. 10, 11

    • Yes → Question 4
    • No → Outcome E
  4. 4

    Is it beyond doubt, for a reasonably well-informed, observant and circumspect person, that they are dealing with AI?

    Art. 50(1) · Guidelines para. 42 to 45

    • Yes, beyond doubt → Outcome B
    • No, or in doubt → Question 5
  5. 5

    Is the system authorised by law to detect, prevent, investigate or prosecute criminal offences, subject to appropriate safeguards, and is it not available to the public to report a criminal offence?

    Art. 50(1), last sentence · Guidelines para. 46 to 49

    • Yes → Outcome C
    • No → Outcome A
  6. 6

    Are you the provider of the system that generates the content, including where it is a general-purpose AI system?

    Art. 50(2) · Guidelines para. 26, 27, 74

    • Yes → Question 7
    • No, we use it → Question 9
  7. 7

    Does the system generate text, images, audio or video that people get to see or hear?

    Art. 50(2) · Guidelines para. 64 to 68

    • Yes → Question 8
    • No → Outcome H
  8. 8

    Does the system do no more than standard editing, such as spelling, formatting, translation or transcription, without changing the meaning, style or intent?

    Art. 50(2) · Guidelines para. 90 to 92

    • Yes, standard editing only → Outcome G
    • No, it creates new content or changes the meaning → Outcome F
  9. 9

    Do you, in a professional capacity, use AI to generate or manipulate image, audio or video that appreciably resembles real or plausible persons, animals, objects, places or events, and that your audience could falsely take to be authentic?

    Art. 3(60) · Art. 50(4) · Guidelines para. 112 to 116

    • Yes → Question 10
    • No → Question 11
  10. 10

    Does your audience immediately see that it is part of an artistic, creative, satirical or fictional work, and is it not news, public information or product advertising?

    Art. 50(4) · Guidelines para. 119 to 124

    • Yes, evidently → Outcome J
    • No, or in doubt → Outcome I
  11. 11

    Do you publish AI-generated or manipulated text that an indeterminate, fairly large group of readers can access, including behind a subscription?

    Art. 50(4), second subparagraph · Guidelines para. 131

    • Yes → Question 12
    • No → Outcome M
  12. 12

    Does the text seek to convey knowledge, facts or opinions?

    Art. 50(4), second subparagraph · Guidelines para. 131

    • Yes → Question 13
    • No → Outcome M
  13. 13

    Is the text about a matter of public interest, such as politics, government, justice, fundamental rights, security, health, the environment, consumer safety or an economic, scientific or cultural development?

    Art. 50(4), second subparagraph · Guidelines para. 131

    • Yes → Question 14
    • No → Outcome M
  14. 14

    Did someone with knowledge of the subject check the content and facts before publication, with no substantive AI changes afterwards?

    Art. 50(4), second subparagraph · Guidelines para. 134 to 136

    • Yes → Question 15
    • No → Outcome K
  15. 15

    Does a person or organisation hold editorial responsibility, and is it public who that is, with contact details?

    Art. 50(4), second subparagraph · Guidelines para. 137, 138

    • Yes → Outcome L
    • No → Outcome K
  16. 16

    Does the system infer emotions, intentions or characteristics from face, voice, body or behaviour?

    Art. 3(39) and (40) · Art. 50(3) · Guidelines para. 101, 103

    • Yes → Question 17
    • No → Outcome P
  17. 17

    Do you use emotion recognition in the workplace or in education, without a medical or safety reason?

    Art. 5(1)(f)

    • Yes → Outcome Q
    • No → Question 18
  18. 18

    Does the system infer race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation from biometric data?

    Art. 5(1)(g)

    • Yes → Outcome R
    • No → Question 19
  19. 19

    Is the system permitted by law to detect, prevent or investigate criminal offences, subject to appropriate safeguards?

    Art. 50(3) · Guidelines para. 109

    • Yes → Outcome O
    • No → Outcome N

Outcome A · Applies

Paragraph 1 applies: disclose that people are dealing with AI

Design the system so that people know they are communicating with AI, at the latest at the first interaction, in a clear, distinguishable and accessible way (paragraph 5). A sentence in the terms or the word ‘assistant’ is not enough (Guidelines para. 38).

Art. 50(1) · Art. 50(5) · Guidelines para. 30 to 45

Outcome B · Does not apply

No separate notice needed, provided you can justify it

Where it is beyond doubt for a reasonably well-informed, observant and circumspect person that they are dealing with AI, the provider need not disclose it separately. You must be able to demonstrate that assessment. Consumer law may still require you to disclose the use of AI (Guidelines para. 50).

Art. 50(1) · Guidelines para. 42 to 45 · Guidelines para. 50

Outcome C · Does not apply

Law enforcement exception

The disclosure obligation of paragraph 1 does not apply to systems authorised by law for law enforcement, subject to appropriate safeguards, as long as the public does not use them to report a criminal offence. The test applies per use: for other uses the obligation applies again (Guidelines para. 48).

Art. 50(1) · Guidelines para. 46 to 49

Outcome D · Does not apply

Paragraph 1 does not apply to this function

There is no direct interaction between a person and AI. If the system does generate content, check paragraphs 2 and 4 via ‘Creates text, images, audio or video’.

Art. 50(1) · Guidelines para. 30

Outcome E · Check further

Paragraph 1 rests on the provider: make sure the notice stays

As a deployer you do not carry the paragraph 1 obligation yourself. The provider must build in the notice; you make sure your configuration or styling does not remove it. If you substantially modify the system or have it developed for you, and put it into service under your own name, you become the provider (Guidelines para. 11).

Art. 3(3) · Art. 3(4) · Art. 50(1) · Guidelines para. 10 to 16

Outcome F · Applies

Paragraph 2 applies: mark the output in a machine-readable way

The output must be marked in a machine-readable format and be detectable as artificially generated, with solutions that are effective, interoperable, robust and reliable as far as technically feasible. ‘Technically feasible’ is objective: what available technology allows (Guidelines para. 81). This applies since 2 August 2026; if the system was on the market before that date, from 2 December 2026 at the latest. That transition only covers the marking, not the chatbot disclosure (Guidelines para. 153).

Art. 50(2) · Guidelines para. 69 to 92 · Guidelines para. 153

Outcome G · Does not apply

No marking obligation for this output

Standard editing that does not change meaning, style or intent need not be marked: spelling, light style polishing, translation, formatting, noise reduction, light cropping, blurring faces, transcription and assistive technology for persons with disabilities (Guidelines para. 92).

Art. 50(2) · Guidelines para. 90 to 92

Outcome H · Does not apply

No marking obligation for this output

Single words, captions, alt text, button labels, source code, machine-to-machine messages and intermediate steps in a closed production workflow fall outside paragraph 2. Only the final output that people see or hear counts.

Art. 50(2) · Guidelines para. 64 to 68

Outcome I · Applies

Paragraph 4 applies: disclose that it is a deep fake

Disclose that the content has been artificially generated or manipulated, at the latest at first exposure (paragraph 5). The tool's watermark does not count: the label must be visible or audible without any aid (Guidelines para. 117). Only use authorised by law for law enforcement is exempt.

Art. 50(4) · Art. 50(5) · Art. 3(60) · Guidelines para. 112 to 118

Outcome J · Partly applies

Paragraph 4 applies in a lighter form

For evidently artistic, creative, satirical or fictional work, you disclose the existence of the generated or manipulated content in a way that does not hamper the display or enjoyment of the work. It is a lighter obligation, not an exemption (Guidelines para. 123). Image rights, the GDPR and copyright continue to apply (Guidelines para. 124).

Art. 50(4) · Guidelines para. 119 to 124

Outcome K · Applies

Paragraph 4 applies: disclose that the text was made with AI

You publish AI-generated or manipulated text to inform the public on a matter of public interest, without substantive human review or without someone holding editorial responsibility. You must then disclose that the text was generated or manipulated with AI.

Art. 50(4), second subparagraph · Guidelines para. 130 to 143 · Guidelines para. 154

Outcome L · Does not apply

No disclosure needed for this text

With substantive human review before publication and a publicly identified party holding editorial responsibility, the text obligation of paragraph 4 does not apply. The exception only holds if you actually follow that way of working.

Art. 50(4), second subparagraph · Guidelines para. 133 to 140

Outcome M · Does not apply

Article 50 requires no disclosure from you for this content

As a deployer you have no disclosure obligation of your own under Article 50 for this content. The provider of the system still marks the output in a machine-readable way (paragraph 2). Consumer law, the GDPR and copyright continue to apply.

Art. 50(2) · Art. 50(4) · Guidelines para. 131

Outcome N · Applies

Paragraph 3 applies: inform the people exposed to it

This obligation rests on the deployer and also applies to processing after the fact (Guidelines para. 100). Inform people about the operation of the system, at the latest at first exposure. Emotion recognition is listed in Annex III and, according to the guidelines, is high-risk where not prohibited (Guidelines para. 102); those requirements apply from 2 December 2027. Biometric categorisation falls under paragraph 3 even when not high-risk, for example age or gender estimation (Guidelines para. 104). Process the data in line with the GDPR; biometric data are often special category data.

Art. 50(3) · Guidelines para. 99 to 110

Outcome O · Does not apply

Law enforcement exception

Paragraph 3 does not apply to systems permitted by law to detect, prevent or investigate criminal offences, subject to appropriate safeguards and in accordance with Union law.

Art. 50(3) · Guidelines para. 109

Outcome P · Does not apply

Paragraph 3 does not apply to this function

Without biometric data there is no emotion recognition or biometric categorisation within the meaning of the AI Act. If you analyse, for example, the sentiment of written messages, the GDPR still applies.

Art. 3(39) and (40) · Guidelines para. 101, 103

Outcome Q · Prohibited

Prohibited: emotion recognition at work or school

Emotion recognition in the workplace and in education is prohibited, except for medical or safety reasons. Informing people does not make it permissible. Pure fatigue detection that infers no emotions falls outside the definition.

Art. 5(1)(f) · Art. 3(39)

Outcome R · Prohibited

Prohibited: biometric categorisation on sensitive traits

Categorising people on the basis of biometric data to infer these traits is prohibited. There is a narrow exception for labelling or filtering lawfully acquired biometric datasets and for law enforcement.

Art. 5(1)(g)

A decision tree simplifies. If you are unsure at a question, read the provision or have it assessed. Checked against the legal text on 5 October 2026.

Open the full decision tree: per outcome the steps from the guidelines, the evidence and examples of good notices

Article 50 explained

What does Article 50 of the AI Act regulate?

Article 50 of the AI Act requires organisations to let people know when they are dealing with AI. It contains four duties: a chatbot or other system for direct interaction makes clear that it is AI (paragraph 1), AI-generated audio, image, video and text is marked in a machine-readable way (paragraph 2), whoever uses emotion recognition or biometric categorisation informs the people concerned (paragraph 3), and deep fakes and AI text on matters of public interest are disclosed as artificial (paragraph 4).

Paragraphs 1 and 2 fall to the provider, paragraphs 3 and 4 to the deployer. Paragraph 5 sets how you inform people for all four: clearly and distinguishably, at the latest at the first interaction or exposure, and accessible to people with disabilities. Article 50 has applied since 2 August 2026, also when a system is not high-risk.

Article 50 is not a high-risk regime: there is no conformity assessment, no Annex IV technical documentation and no registration in the EU database. Nor does it mean that all AI content needs a visible label. A visible disclosure is required for deep fakes and for AI text on matters of public interest; the marking in paragraph 2 is machine-readable.

Who has to do what under Article 50?

The four situations side by side, with the role that carries the duty. One organisation can hold a different role for each system.

ParagraphSituationWho carries the dutyWhat you doException
Para. 1People communicate directly with AI, such as a chatbot, voicebot or AI assistantProviderDesign the system so that people know they are communicating with AI, at the latest at the first interactionWhere it is obvious to a reasonably well-informed, observant person; law enforcement authorised by law, unless the public can use it to report a criminal offence
Para. 2AI generates synthetic audio, image, video or textProvider, including of general-purpose AI systemsMark the output in a machine-readable format and make it detectable as artificially generated or manipulatedAn assistive function for standard editing, or no substantial change to the input; law enforcement authorised by law
Para. 3Emotion recognition or biometric categorisationDeployerInform the people exposed about the operation of the system and process personal data in line with the GDPRLaw enforcement authorised by law. First check the ban on emotion recognition at work and in education (Article 5(1)(f))
Para. 4Deep fakes, and AI text that informs the public on matters of public interestDeployerDisclose that the content has been artificially generated or manipulatedEvidently artistic, creative, satirical or fictional work: a lighter disclosure that does not hamper the work. Text: no disclosure after human review or editorial control under editorial responsibility

If you build a chatbot on someone else's language model and put it into service under your own name, you are the provider. If you only place a supplier's chatbot on your site, you are the deployer: you do not carry the paragraph 1 duty yourself, but your configuration must not remove the disclosure (guidelines, points 11 and 12).

Article 50 does not affect the requirements for high-risk systems or other transparency obligations in Union or national law, such as consumer law (paragraph 6). If your system is also high-risk, those requirements apply as well.

Since when does Article 50 apply?

Article 50 has applied since 2 August 2026. The Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026, did not move that date. It did move the high-risk obligations: Annex III to 2 December 2027 and Annex I to 2 August 2028.

There is one transition. Providers of generative systems that were already on the market before 2 August 2026 have until 2 December 2026 for the machine-readable marking in paragraph 2. That transition does not apply to the chatbot disclosure or to paragraphs 3 and 4 (guidelines, point 153). A system placed on the market after 2 August 2026 marks from day one.

Since 2 August 2026, national supervisory authorities enforce Article 50. An infringement can lead to a fine of up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher; for SMEs, start-ups and small mid-caps the lower of the two applies (Article 99(4), (6) and (6a)).

What does Article 50 look like in practice?

A customer service chatbot on your website

If you buy the chatbot in, the supplier is the provider and delivers the AI disclosure. You check that it is there and do not remove it through your configuration. A short sentence at the start is enough, for example: "You are chatting with a virtual assistant based on AI. Want to speak to an employee? Type employee." A human name, an avatar or just the word "assistant" is not enough (guidelines, point 38). For a voicebot, the disclosure is audible at the start of the call.

A supplier of an image generator for business customers

The supplier is the provider and, under paragraph 2, ensures that every generated image is marked in a machine-readable way and detectable. If a customer uses those images in a campaign and an image is a deep fake, that customer has its own disclosure duty as deployer under paragraph 4.

A municipality publishes an AI-generated explainer video

The municipality is the deployer. If the video resembles existing persons, places or events and could falsely appear authentic, it is a deep fake (Article 3(60)) and the municipality discloses that the content is artificial. An animation without that appearance of authenticity does not need a label under Article 50. If an agency makes the video under the municipality's direction, the disclosure duty stays with the municipality (guidelines, point 14).

A newsroom uses AI for an article on a public debate

Paragraph 4 then requires a disclosure, unless a person has reviewed the text and a natural or legal person holds editorial responsibility. Record in an editorial policy who reviews and who is responsible, and name that person publicly, for example in the masthead (guidelines, points 134 to 138).

A contact centre analyses emotions in calls

Emotion recognition of employees at work is prohibited, except for medical or safety reasons (Article 5(1)(f)). For customers that ban does not apply: there you inform them about the operation of the system (paragraph 3), and from 2 December 2027 such a system is high-risk (Annex III, point 1(c)). Sentiment analysis of the text of a call alone is not emotion recognition within Article 3(39), because that must be based on biometric data.

Where does it usually go wrong?

  • Folding the provider and deployer duties into one vague labelling task, so supplier and customer point at each other.
  • Hiding the disclosure in the terms and conditions, a footer or metadata nobody sees. Paragraph 5 requires a clear disclosure at the moment of first contact.
  • Thinking Article 50 is only about deep fakes. It also covers chatbots, all synthetic output, and emotion recognition and biometric categorisation.
  • Relying on the editorial control exception without demonstrable review and an identifiable responsible person.

Which code of practice and guidelines belong to Article 50?

The European Commission adopted its guidelines on the transparency obligations of Article 50 on 20 July 2026 (C(2026) 5054 final). They explain, among other things, when AI interaction is obvious, who is provider or deployer and how narrow the exception for artistic work is: advertising does not automatically fall under it (guidelines, point 122).

The Code of Practice on Transparency of AI-generated Content was published on 10 June 2026. It is voluntary and has a section for providers (marking and detection) and one for deployers (deep fakes and AI text on matters of public interest). Following a positive adequacy assessment, signatories can use it to support compliance with Article 50(2), (4) and (5); paragraph 3 is not covered. Those who do not sign substantiate their own measures (guidelines, point 148).

In depth per paragraph

Have your own situation checked

Walk through the decision tree for your role and use case, or put your own case to the case check. If you want it reviewed, Praxikon assesses your chatbots, voicebots and AI content against Article 50, with the reasoning and the disclosure text for each use case.

Article 50 decision treeCheck your caseArt. 50 checkAbout Zahed Ashkara

Sources

Updated on 10 October 2026

Official guidance on this article

13

What does this mean for you?

Provider+
You are the provider if you develop the AI system or have it developed and place it on the market or put it into service under your own name. Under paragraph 1 you design a chatbot or other system for direct interaction so that people know they are communicating with AI, unless that is obvious. Under paragraph 2 you mark generated audio, image, video and text in a machine-readable and detectable way; for systems already on the market before 2 August 2026 that marking runs until 2 December 2026 at the latest. Give deployers information on limitations and exceptions so they can meet their own duties.
Deployer+
You are the deployer if you use an AI system under your own authority. Under paragraph 3 you inform people exposed to emotion recognition or biometric categorisation. Under paragraph 4 you disclose that a deep fake or AI text on matters of public interest is artificial, unless an exception applies. If you use a chatbot you bought in, your configuration does not remove the provider's AI disclosure.
SME / Startup+
Article 50 has no exemption for small organisations. If you use a chatbot you bought in on your website, check that it discloses at first contact that it is AI. If you publish AI images or audio that falsely appear to show existing persons, places or events, disclose that they are artificial. For fines, SMEs and start-ups face the lower of the two maximum amounts (Article 99(6)).
Public Sector+
Public bodies have the same duties, without a separate transition. A chatbot on a municipal website must disclose that it is AI, an AI-generated public information video that falsely appears authentic gets a disclosure, and AI text on matters of public interest gets a disclosure unless a person has reviewed it editorially and the organisation holds responsibility for it. The exceptions only cover detection and prosecution of criminal offences authorised by law.

Article 50 checklist

  • Map for each system whether it communicates directly with people, generates content, recognises emotions or categorises biometrically, or produces deep fakes or public-facing text.
  • Determine your role for each system: provider (paragraphs 1 and 2), deployer (paragraphs 3 and 4) or both.
  • For emotion recognition, first check the ban in Article 5(1)(f) (work and education) before relying on the information duty.
  • Show the AI disclosure of a chatbot or voicebot at first contact, visibly or audibly, and not only in the terms.
  • Mark generated output in a machine-readable way, for example with C2PA provenance data and a watermark, and test whether the marking survives compression, cropping and conversion.
  • Disclose deep fakes visibly or audibly and record, for AI text on matters of public interest, who reviews it editorially and who is responsible.
  • Agree in supplier contracts who delivers the marking and the AI disclosure, also after model updates.
  • Keep evidence: screenshots of disclosures, the chosen marking standard, test results and the reasoning behind each exception relied on.

Related recitals

Recital 132Primary

Certain AI systems intended to interact with natural persons or to generate content may pose specific risks of impersonation or deception irrespective of whether they qualify as high-risk or not. In c…

Recital 133Primary

A variety of AI systems can generate large quantities of synthetic content that becomes increasingly hard for humans to distinguish from human-generated and authentic content. The wide availability an…

134

Related tools

AI Decision Tree

Related enforcement

  • •
    Italy fines OpenAI €15 million for ChatGPT privacy violationsGarante per la protezione dei dati personali · Dec 2024↗
  • •
    Italy investigates ChatGPT — temporary ban lifted after measuresGarante per la protezione dei dati personali · May 2023↗

Related blog posts

Cross-references

Articles

Art. 6Art. 52

Recitals

132133134

Frequently asked questions

What does Article 50 of the AI Act regulate?+
Article 50 requires transparency about AI in four situations: a system for direct interaction, such as a chatbot, discloses that it is AI (paragraph 1); generated audio, image, video and text is marked in a machine-readable way (paragraph 2); whoever uses emotion recognition or biometric categorisation informs the people concerned (paragraph 3); and deep fakes and AI text on matters of public interest are disclosed as artificial (paragraph 4).
When do the transparency obligations of Article 50 take effect?+
They have applied since 2 August 2026. The Digital Omnibus (Regulation (EU) 2026/1744) did not move that date. Only for the machine-readable marking in paragraph 2 do providers of systems already on the market before 2 August 2026 have until 2 December 2026; that transition does not apply to the chatbot disclosure (guidelines, point 153).
Which duties fall to the provider and which to the deployer?+
Paragraph 1 (disclosure of direct AI interaction) and paragraph 2 (machine-readable marking) fall to the provider. Paragraph 3 (emotion recognition and biometric categorisation) and paragraph 4 (deep fakes and AI text on matters of public interest) fall to the deployer. One organisation can hold a different role for each system.
Do I need to tell users they are talking to a chatbot?+
The provider must design a chatbot so that users know they are communicating with AI, at the latest at the first interaction, unless that is obvious to a reasonably well-informed and observant person. A human name, an avatar or just the word 'assistant' is not enough (guidelines, point 38). If you use a supplier's chatbot, make sure your configuration does not remove that disclosure.
Does all AI-generated content need a visible label?+
No. A visible disclosure is required for deep fakes and for AI text published to inform the public on matters of public interest (paragraph 4). The marking the provider applies under paragraph 2 is machine-readable and does not have to be visible.
Do I need to label deepfakes under the AI Act?+
Yes. Article 50(4) requires the deployer who makes or publishes a deep fake to disclose that the image, audio or video content has been artificially generated or manipulated. For evidently artistic, creative, satirical or fictional work, a disclosure that does not hamper the work is enough; advertising does not automatically fall under that (guidelines, point 122).
Does the labelling obligation also apply to AI-generated text on news websites?+
AI text published to inform the public on matters of public interest must be disclosed (paragraph 4). That duty falls away when a person has reviewed the text and a natural or legal person holds editorial responsibility. Record this in an editorial policy; according to the code of practice, a record per text is not required.
What if my AI tool only assists with standard editing?+
Then the paragraph 2 exception may apply: the marking duty does not apply to the extent the system performs an assistive function for standard editing, such as a spelling or grammar checker, or does not substantially alter the input or its meaning (guidelines, point 92).
May I use emotion recognition if I inform people?+
Not always. First check Article 5(1)(f): emotion recognition in the workplace and in education is prohibited, except for medical or safety reasons. A prohibited use is not fixed by a disclosure. If the use falls outside the ban, you inform the people concerned about the operation of the system (paragraph 3).
Does Article 50 apply if my AI system is not high-risk?+
Yes. Article 50 is separate from high-risk classification. If a system is also high-risk, those requirements apply as well: Article 50(6) leaves Chapter III and other transparency obligations in Union or national law unaffected.
Who is responsible if I integrate an external AI model into my service?+
If you build your own system on an external model and put it into service under your own name, you are its provider (guidelines, point 11). Agree in the contract that the model supplier delivers and maintains the machine-readable marking, also after updates, and do not rely on a bare claim of being 'AI Act compliant'.
What evidence should I keep for Article 50?+
Article 50 requires no conformity assessment or technical documentation, but after a complaint a supervisory authority will want to see that the disclosure was there. Keep screenshots of disclosures and labels, the chosen marking standard and test results, the role split per system, the reasoning behind exceptions and the agreements with suppliers.
What is the role of the code of practice and the guidelines for Article 50?+
The Commission adopted guidelines on Article 50 on 20 July 2026. The Code of Practice on Transparency of AI-generated Content of 10 June 2026 is voluntary; following a positive adequacy assessment, signatories can use it to support compliance with paragraphs 2, 4 and 5, not paragraph 3. Those who do not sign substantiate their own measures (guidelines, point 148).
What happens if I do not comply with Article 50?+
Since 2 August 2026, national supervisory authorities enforce Article 50. A fine can reach 15 million euros or 3 percent of worldwide annual turnover, whichever is higher; for SMEs, start-ups and small mid-caps the lower of the two applies (Article 99(4), (6) and (6a)).

What Article 50 requires in practice

Connections

What connects to Article 50 AI Act

Themes where this returns

Where this meets the other law

The counterpart in the other law

GDPR interpretation that also applies here6 of 15