EU AI Act Transparency Obligations: A Reference
This page is written for the person who has to answer “what are our transparency obligations under the AI Act” in a meeting — a structured reference to Article 50's obligations as a set, how they relate to each other, and where enforcement actually comes from, rather than a narrative explainer or a task checklist. For a plainer, less formal version of the same ground, see AI transparency requirements.
The obligations, in order
Disclosure obligation
Article 50(1)A person must be told they're interacting with an AI system, unless that's obvious from context. Applies to chatbots, virtual assistants, emotion recognition, and biometric categorization systems. Owed to the end user, at the point of interaction.
Synthetic content marking obligation
Article 50(2)AI-generated text, image, audio, and video output must be marked in a machine-readable, technically detectable format — not just visibly labeled. This is a technical requirement about the content itself, not about what a human sees on the page around it.
Published content labeling obligation
Article 50(4)AI-generated text published to inform the public needs a human-visible label, unless a person took substantive editorial responsibility for it. AI-generated or manipulated image, audio, and video content (including deepfakes) must be labeled as such.
Evidentiary obligation
Implied by enforceabilityNot a separately numbered duty, but a practical one: an obligation that can't be demonstrated after the fact isn't compliance, it's an assertion. Regulators and enforcement bodies need a way to verify disclosure and labeling actually happened, which in practice means an audit trail.
How the obligations interact
These aren't four independent checkboxes — they compound. A chatbot that generates and publishes content, for instance, carries the disclosure obligation for the conversation itself and a separate labeling obligation for anything it generates that gets published externally. Satisfying one doesn't imply the other is covered, and a system can be fully disclosure-compliant while still failing the content-labeling obligation for what it produces.
The evidentiary obligation sits underneath all three: it's what turns “we do this” into something a regulator, a customer's legal team, or an internal audit can actually verify. Without it, the other three obligations are only as credible as the word of whoever built the system.
Enforcement and penalties
Article 50 has been enforceable since 2 August 2026. Penalties can reach €15 million or 3% of global annual turnover, whichever is higher — the same enforcement tier as GDPR, even though the underlying obligations here are narrower in scope. For SMEs and start-ups, the AI Act applies the lower of the two amounts (Art. 99(6)). Enforcement runs through each member state's designated market surveillance authority, not a single central EU regulator.
Meeting all four obligations at once
Disclarion's SDK is built around this exact structure — a disclosure mechanism, content labeling, and an insert-only audit trail, all produced from the same call that logs an interaction, across OpenAI, Anthropic, and Gemini. The obligations map directly onto product features rather than requiring four separate systems bolted together.
Frequently asked questions
Are these four separate legal obligations or one combined rule?
Three are directly specified in Article 50's text (disclosure, synthetic content marking, published content labeling). The fourth — being able to demonstrate compliance — follows from the practical reality of enforcement rather than being its own numbered clause, but functions as an obligation in practice.
Who enforces Article 50 obligations?
Enforcement runs through the national market surveillance authorities each EU member state designates under the AI Act, with penalties that can reach €15 million or 3% of global annual turnover, whichever is higher. For SMEs and start-ups, the AI Act applies the lower of the two amounts (Art. 99(6)).
Do all four obligations apply to every AI system?
The disclosure obligation applies to interactive and analytical systems (chatbots, emotion recognition, biometric categorization). The content-marking and labeling obligations apply specifically to systems that generate synthetic content. Most products fall under a subset, not all four at once.
Is meeting these obligations the same as being 'AI Act compliant' overall?
No. These are the Article 50 transparency obligations specifically. The AI Act separately covers high-risk system requirements (Annex III, generally applying from December 2027) and prohibited-practice rules, which are different obligations entirely.