EU AI Act Article 50: Transparency Obligations, Implemented in Practice
EU AI Act Article 50 requires that people are told when they are talking to an AI, that synthetic content is marked in a machine-readable way, and that deep fakes and emotion recognition are disclosed. At SIFAMO we turn that into a response rule in the gateway: no disclosure state, no answer; no provenance record, no image.
Article 50 is the part of the AI Act that hits most companies first, long before the high-risk obligations. Every bank has a chatbot, every marketing team generates images and text with generative models, many contact-centre platforms offer sentiment detection. Article 50 of the AI Regulation addresses providers (paragraphs 1 and 2) and deployers (paragraphs 3 and 4) of such systems, regardless of risk class.
The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. The Omnibus did not move the date for Article 50. The overview is in EU AI Act in practice.
What Article 50 requires
Art. 50(1): Providers of AI systems intended to interact directly with natural persons design them so that the person is informed that they are interacting with an AI. Excepted: cases where this is obvious to a reasonably well-informed and observant person, and systems authorised by law for law enforcement, unless publicly available for reporting offences.
Art. 50(2): Providers of systems, including GPAI systems, that generate synthetic audio, image, video or text content mark the outputs in a machine-readable format as artificially generated or manipulated, effective, interoperable, robust and reliable as far as technically feasible. Excepted: assistive standard editing, no substantial alteration of the input data, law enforcement. Recital 133 names watermarks, metadata, cryptographic provenance methods, logging and fingerprints.
Art. 50(3): Deployers of an emotion recognition system or a biometric categorisation system (Art. 3(39) and (40)) inform the exposed persons about the operation of the system and comply with the GDPR. Excepted: law enforcement.
Art. 50(4), first sentence: Deployers who generate or manipulate deep fakes under Art. 3(60) disclose this. Excepted: law enforcement; for evidently artistic, creative, satirical or fictional works a disclosure that does not hamper the enjoyment of the work is enough.
Art. 50(4), second sentence: Deployers who publish AI-generated or manipulated text to inform the public on matters of public interest disclose this. Excepted: law enforcement and text with human review or editorial control, where a person or organisation holds editorial responsibility.
Art. 50(5) and (6): Clear and distinguishable, at the latest at the first interaction or exposure, accessible. Chapter III and other transparency obligations under Union or national law remain untouched.
Art. 50(7): The Commission encourages codes of practice at Union level for detecting and labelling artificially generated content and assesses them under Art. 56(6); if a code is not adequate, it can adopt an implementing act under Art. 98(2).
The technical example
A bank runs a customer-service chatbot on its website, in its app and on the phone. The bank built it on a hosted model and puts it into service under its own name, so it is the provider under Art. 3(3) and the deployer at the same time. Every message passes with a session ID through the policy gateway, which pseudonymises, checks the registry and calls the model. Art. 50(1) bites before the first answer, hardest on the phone, because a natural voice creates the impression of a human.
The marketing team of the same bank uses a GPAI service to generate campaign images and draft texts: photorealistic scenes with people, stylised illustrations, social media posts, explainers on interest-rate trends. Paragraph 2 hits the provider of the service, paragraph 4 the bank as deployer: a photorealistic image of an advisor who does not exist is a deep fake under Art. 3(60). A text on interest-rate trends can inform the public on a matter of public interest and needs the disclosure too, unless an editor takes responsibility.
Third, the contact-centre platform offers an add-on: sentiment detection from the voice, for customers and, as “coaching”, for employees. From the voice means from biometric data, so this is an emotion recognition system under Art. 3(39). For customers paragraph 3 applies, for employees the prohibition under Art. 5(1)(f). The text sentiment the chatbot derives from written messages does not fall under it, because text is not biometric data.
How we implement it
At SIFAMO we attach all four obligations to the one place that sees every request anyway: the policy gateway in the governance layer. Transparency in the prompt is a request to the model. Transparency the gateway enforces is a property of the system.
A registry field for the Art. 50 class. Every registry record has a field art50_class with the classes: interacts with persons, generates synthetic content, emotion recognition, biometric categorisation, deep-fake capable, public-interest text, none. The field is set in the first gate and countersigned by risk; “none” needs a justification, “emotion recognition” the notice text and the reference to the data protection impact assessment. The gateway loads the response policy from the class. No class, no route.
Disclosure as a response policy. For the chatbot, the gateway keeps a disclosure state per session, empty at the start. The channel sets it as soon as the notice has been rendered or spoken, with text version, language and accessibility variant. If the state is empty, the gateway returns no model answer but the notice itself. On top of that, an output rule: if someone asks “am I talking to a human”, a classifier checks the answer and replaces any denial of its own nature with the standard response. We do not use the “obvious” exception. The evidence log stores notice version, timestamp and channel per session.
Marking as an output stage. Images, audio and video from the gateway pass an output stage before the asset reaches the application. It checks the model provider’s watermark and adds a signed provenance record to the metadata: registry ID of the generating system, model and version, timestamp, hash of the prompt, flag “artificially generated”. Hash and record ID go into the evidence log, so that an asset stays traceable even without metadata. For text we are honest: watermarks are not robust, metadata does not survive copy and paste. We store the hash of the text in the log and rely on the visible labelling under paragraph 4.
Deep fakes and editorial control in the CMS. Before publication the marketing system queries the registry. If the source is deep-fake capable, it requires a visible label for photorealistic people and voices; for stylised illustrations the note in the asset details is enough, our reading of the exception for creative works. For texts on interest rates, markets or regulation, the workflow requires the signature of a named editor; then the editorial-responsibility exception applies, and the signature is in the log. Without a signature the label stays. We also mark where paragraph 2 strictly hits the service provider, because nobody in marketing should have to settle the bank’s role tool by tool.
Emotion recognition with a data zone. The sentiment module is registered as its own system, class emotion recognition, scope: customers only. On every call the gateway checks the data zone of the speaker identifier; identifiers from the employee directory are the red zone for this system, the call is rejected and counted. The vendor’s coaching feature is switched off, and a configuration test checks daily that it stays off. The notice to customers sits in the greeting of the voice dialogue, before the first call to the module.
Monitoring for unmarked outputs. The first monitor counts model answers without a disclosure state; the number has to be zero. The second samples published assets on the website and in the social channels and checks watermark and provenance record. An asset with a hash in the log but no record raises an alert to the head of marketing; usually the CDN stripped the metadata while resizing. An asset that looks generated and is in no log is shadow generation and raises a ticket. The third checks that the sentiment module has never seen an employee identifier. The auditor sees the disclosure record per session, provenance record and labelling decision per asset, the class with signature per system, plus the monitor reports. I am an engineer, not a lawyer. But a transparency obligation that exists only in the policy document is gone by the ten-thousandth session. One the gateway enforces is still there.
Typical mistakes
“Everyone knows it is a bot.” The “obvious” exception does not hold on the phone or in a chat with a name and an avatar. A sentence in the terms of use does not replace a disclosure at the first interaction.
Disclosure in the system prompt. “Say you are an AI if asked” is an instruction the model forgets under pressure.
Trusting the provider’s watermark. The visible disclosure under paragraph 4 stays with the deployer, and metadata rarely survives your own image pipeline.
Confusing sentiment with emotion, in both directions. Text sentiment is not emotion recognition under Art. 3(39); voice analysis is. The coaching feature for employees is prohibited under Art. 5, not merely subject to disclosure.
EU AI Act Article 50 checklist
Art. 50 class set in the registry for every system, countersigned by risk, “none” justified.
Chatbots and voice dialogues return no model answer without a disclosure state; notice text versioned and accessible; output rule against denial of the system’s own AI nature.
Generated images, audio and video carry a watermark and a signed provenance record; hash and record in the evidence log.
Generated text hashed into the log; visible label or documented editorial responsibility for public-interest text.
Deep-fake capable sources trigger a visible label in the CMS; the exception for creative works is documented.
Emotion recognition: notice before the first call, DPIA linked, employee identifiers in the red zone, vendor feature off and tested daily.
Monitoring for unmarked and for unregistered outputs, with alerts to the business owner.
Open questions
What “matters of public interest” means for corporate content is open. That a bank’s market commentary counts and a product text does not is my reading, not a settled one.
Whether watermark plus signed provenance record counts as state of the art under paragraph 2 will be decided with the code of practice under paragraph 7 and the Commission’s assessment of it.
Frequently asked questions
Does this apply to us as a deployer?
Yes, directly. Paragraphs 3 and 4 address the deployer: notice for emotion recognition, disclosure for deep fakes and for public-interest text. Paragraphs 1 and 2 address the provider, but whoever builds the chatbot and runs it under their own name is a provider too. And with a purchased chatbot I would check whether the notice appears in your channel. The basics are in AI Governance.
Do we have to label every AI-generated text?
No. The machine-readable marking under paragraph 2 is the duty of the provider of the generating system. A visible label is required by paragraph 4 only for text that informs the public on matters of public interest, and not even there under editorial responsibility. An internal draft needs no label, but it needs a log entry.
May we use sentiment detection in the contact centre?
For customers yes, with a notice before first use and a GDPR basis under paragraph 3. For employees no: Art. 5(1)(f) prohibits emotion recognition in the workplace, except for medical or safety reasons, since February 2025. That is a prohibition, not a transparency matter. The other articles in the series: all articles at a glance.
EU AI Act Article 50: Transparency Obligations, Implemented in Practice
EU AI Act Article 50 requires that people are told when they are talking to an AI, that synthetic content is marked in a machine-readable way, and that deep fakes and emotion recognition are disclosed. At SIFAMO we turn that into a response rule in the gateway: no disclosure state, no answer; no provenance record, no image.
Article 50 is the part of the AI Act that hits most companies first, long before the high-risk obligations. Every bank has a chatbot, every marketing team generates images and text with generative models, many contact-centre platforms offer sentiment detection. Article 50 of the AI Regulation addresses providers (paragraphs 1 and 2) and deployers (paragraphs 3 and 4) of such systems, regardless of risk class.
The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. The Omnibus did not move the date for Article 50. The overview is in EU AI Act in practice.
What Article 50 requires
The technical example
A bank runs a customer-service chatbot on its website, in its app and on the phone. The bank built it on a hosted model and puts it into service under its own name, so it is the provider under Art. 3(3) and the deployer at the same time. Every message passes with a session ID through the policy gateway, which pseudonymises, checks the registry and calls the model. Art. 50(1) bites before the first answer, hardest on the phone, because a natural voice creates the impression of a human.
The marketing team of the same bank uses a GPAI service to generate campaign images and draft texts: photorealistic scenes with people, stylised illustrations, social media posts, explainers on interest-rate trends. Paragraph 2 hits the provider of the service, paragraph 4 the bank as deployer: a photorealistic image of an advisor who does not exist is a deep fake under Art. 3(60). A text on interest-rate trends can inform the public on a matter of public interest and needs the disclosure too, unless an editor takes responsibility.
Third, the contact-centre platform offers an add-on: sentiment detection from the voice, for customers and, as “coaching”, for employees. From the voice means from biometric data, so this is an emotion recognition system under Art. 3(39). For customers paragraph 3 applies, for employees the prohibition under Art. 5(1)(f). The text sentiment the chatbot derives from written messages does not fall under it, because text is not biometric data.
How we implement it
At SIFAMO we attach all four obligations to the one place that sees every request anyway: the policy gateway in the governance layer. Transparency in the prompt is a request to the model. Transparency the gateway enforces is a property of the system.
A registry field for the Art. 50 class. Every registry record has a field art50_class with the classes: interacts with persons, generates synthetic content, emotion recognition, biometric categorisation, deep-fake capable, public-interest text, none. The field is set in the first gate and countersigned by risk; “none” needs a justification, “emotion recognition” the notice text and the reference to the data protection impact assessment. The gateway loads the response policy from the class. No class, no route.
Disclosure as a response policy. For the chatbot, the gateway keeps a disclosure state per session, empty at the start. The channel sets it as soon as the notice has been rendered or spoken, with text version, language and accessibility variant. If the state is empty, the gateway returns no model answer but the notice itself. On top of that, an output rule: if someone asks “am I talking to a human”, a classifier checks the answer and replaces any denial of its own nature with the standard response. We do not use the “obvious” exception. The evidence log stores notice version, timestamp and channel per session.
Marking as an output stage. Images, audio and video from the gateway pass an output stage before the asset reaches the application. It checks the model provider’s watermark and adds a signed provenance record to the metadata: registry ID of the generating system, model and version, timestamp, hash of the prompt, flag “artificially generated”. Hash and record ID go into the evidence log, so that an asset stays traceable even without metadata. For text we are honest: watermarks are not robust, metadata does not survive copy and paste. We store the hash of the text in the log and rely on the visible labelling under paragraph 4.
Deep fakes and editorial control in the CMS. Before publication the marketing system queries the registry. If the source is deep-fake capable, it requires a visible label for photorealistic people and voices; for stylised illustrations the note in the asset details is enough, our reading of the exception for creative works. For texts on interest rates, markets or regulation, the workflow requires the signature of a named editor; then the editorial-responsibility exception applies, and the signature is in the log. Without a signature the label stays. We also mark where paragraph 2 strictly hits the service provider, because nobody in marketing should have to settle the bank’s role tool by tool.
Emotion recognition with a data zone. The sentiment module is registered as its own system, class emotion recognition, scope: customers only. On every call the gateway checks the data zone of the speaker identifier; identifiers from the employee directory are the red zone for this system, the call is rejected and counted. The vendor’s coaching feature is switched off, and a configuration test checks daily that it stays off. The notice to customers sits in the greeting of the voice dialogue, before the first call to the module.
Monitoring for unmarked outputs. The first monitor counts model answers without a disclosure state; the number has to be zero. The second samples published assets on the website and in the social channels and checks watermark and provenance record. An asset with a hash in the log but no record raises an alert to the head of marketing; usually the CDN stripped the metadata while resizing. An asset that looks generated and is in no log is shadow generation and raises a ticket. The third checks that the sentiment module has never seen an employee identifier. The auditor sees the disclosure record per session, provenance record and labelling decision per asset, the class with signature per system, plus the monitor reports. I am an engineer, not a lawyer. But a transparency obligation that exists only in the policy document is gone by the ten-thousandth session. One the gateway enforces is still there.
Typical mistakes
EU AI Act Article 50 checklist
Open questions
Frequently asked questions
Does this apply to us as a deployer?
Yes, directly. Paragraphs 3 and 4 address the deployer: notice for emotion recognition, disclosure for deep fakes and for public-interest text. Paragraphs 1 and 2 address the provider, but whoever builds the chatbot and runs it under their own name is a provider too. And with a purchased chatbot I would check whether the notice appears in your channel. The basics are in AI Governance.
Do we have to label every AI-generated text?
No. The machine-readable marking under paragraph 2 is the duty of the provider of the generating system. A visible label is required by paragraph 4 only for text that informs the public on matters of public interest, and not even there under editorial responsibility. An internal draft needs no label, but it needs a log entry.
May we use sentiment detection in the contact centre?
For customers yes, with a notice before first use and a GDPR basis under paragraph 3. For employees no: Art. 5(1)(f) prohibits emotion recognition in the workplace, except for medical or safety reasons, since February 2025. That is a prohibition, not a transparency matter. The other articles in the series: all articles at a glance.
Related: EU AI Act · All articles · AI Governance · AI Consultant
Recent Posts
Recent Comments