Scenario 1
People interact directly with AI
For chatbots, AI assistants and other direct interactions, Article 50(1) puts the disclosure-by-design duty on the provider, unless the AI nature is obvious in context.
Article 50 AI Act
Transparency obligations for AI systems that are not high-risk
Many organisations do not fall directly under Annex III, but still need to make clear when people interact with AI or when content is generated or manipulated by AI. This is the Article 50 route.
Classify first
Avoid treating a low-risk chatbot as Annex III, or reducing a high-risk use case to a simple disclosure.
Then disclose and document
Link the disclosure to the context: screen text, content label, machine-readable marking, internal instruction and review moment.
Practical guidance for teams
This page is written as practical guidance for legal teams, compliance teams, AI owners and decision-makers.
The four Article 50 scenarios
Article 50 contains several obligations for providers and deployers. Your role determines the evidence and implementation needed.
Scenario 2
AI generates synthetic content
For synthetic audio, image, video or text, Article 50(2) requires the provider to make outputs machine-readable and detectable as artificially generated or manipulated, subject to its exceptions.
Scenario 3
Emotion recognition or biometric categorisation
When people are exposed to emotion recognition or biometric categorisation, the system operation must be disclosed and the GDPR route must be correct.
Scenario 4
Deepfakes and public-facing AI text
Deepfakes and certain AI-generated text published to inform the public need clear disclosure, subject to specific exceptions.
Workflow
Avoid over- or under-classification
- 1First determine whether the system meets the AI Act definition.
- 2Rule out prohibited practices and then check Annex III high-risk scope.
- 3Document whether Article 50 applies separately or alongside another route.
- 4Design the disclosure, marking or notice at the moment people need it.
- 5Keep evidence: system register, screen text, content process, responsible role and periodic review.
Connect this to the right RAIP route
Annex III high-risk
First check whether the system falls within one of the eight high-risk domains.
View Annex IIIArticle 50 legal text
Read Article 50 with internal explanation and cross-references.
Open Article 50AI literacy
Ensure teams know when transparency, human oversight and escalation are required.
View Article 4 evidenceRisk assessment
Document why the system is high-risk, limited-risk or minimal-risk.
Start risk assessmentIn-depth analyses by topic
This page gives the overview. Each obligation is worked out separately, with the legal text, the scope and what you need to be able to demonstrate.
- The 2 August 2026 deadline that was not postponed
- What applies in practice from 2 August 2026
- Checklist: what to do before 2 August
- Provider or deployer: whose obligation is this?
- Labelling and detection in practice
- Making deepfakes recognisable
- Emotion recognition and biometrics: inform or prohibited
- The EU code of practice on AI content transparency
- Enforcement and fines for non-compliance
- Field test: do Dutch chatbots say they are AI?
From knowing to demonstrating
Knowing which obligation applies is step one. Under scrutiny the next question is: on what basis, decided by whom, when, and which facts were on the table. Producing that record is work in its own right, separate from this explanation.
The AI Compliance Agent by Embed AI runs the intake per system, grounds the classification in the legal text and produces a dossier with the basis, the established facts and what is still missing. The demo runs on fictitious data.
Frequently asked questions
Do transparency obligations only apply to high-risk AI?
No. Article 50 sits next to the Annex III high-risk route. An AI system can be outside high-risk scope and still need a clear disclosure, label or detectable marking.
When is the Article 50 route relevant?
The route is relevant for direct interaction with AI, synthetic audio, image, video or text, emotion recognition, biometric categorisation, deepfakes and certain AI-generated text published to inform the public.
Does transparency replace risk classification?
No. Start with classification: prohibited practice, Annex III high-risk, GPAI, Article 50 or minimal obligations. Transparency can then be a separate obligation or part of a wider high-risk route.
What should an organisation document in practice?
Document the AI system, provider and deployer roles, the relevant Article 50 scenario, the disclosure or marking used, when people receive that information and who reviews it periodically.
First determine whether you need Annex III, Article 50 or both routes
Start with a short risk assessment and then document the disclosures, labels, training and governance you actually need.