EU AI Act Article 50: The Complete Guide to Transparency Obligations (2026)

If your organisation puts AI in front of people in the European Union, Article 50 of the EU AI Act asks one deceptively simple thing of you: say so. A chatbot must identify itself as a machine. AI-generated images, audio, video and text must be marked as artificial. Deepfakes and AI-written public-interest content must be labelled. These transparency duties apply from 2 August 2026, and unlike the headline-grabbing rules on high-risk AI, they reach almost every business that touches generative AI — not just a regulated few.

This guide explains exactly what Article 50 of Regulation (EU) 2024/1689 requires: the four obligations, who is bound by each, the exemptions that genuinely apply, how a disclosure must actually be made, the penalties, and a practical path to compliance before the August 2026 deadline.

What Article 50 actually is

The EU AI Act sorts AI into tiers: prohibited practices, high-risk systems, and everything else. Article 50 sits in a category of its own. It imposes transparency obligations that do not depend on a system being high-risk. They apply whenever an AI system is used in one of four specific situations — regardless of how risky the system otherwise is.

That breadth is the whole point. A company with no high-risk AI at all can still carry significant Article 50 duties because it runs a customer-facing chatbot, generates marketing imagery with AI, or publishes AI-assisted articles. Transparency is one of the most common compliance triggers in the entire Act. For many organisations, Article 50 will be their primary AI Act obligation, not a footnote to it.

One important point up front: these duties are cumulative, not alternative. If a high-risk system also interacts directly with people or generates content, both the high-risk obligations (Chapter III) and the Article 50 transparency duties apply at the same time.

The four transparency obligations

The European Commission’s draft guidelines, published on 8 May 2026, describe how the four obligations are intended to apply. Here is each, in plain terms.

1. Interactive AI systems — Article 50(1), the “chatbot rule”

When an AI system is intended to interact directly with people — chatbots, virtual assistants, AI companions, automated phone systems — the provider must design it so users are informed they are dealing with an AI rather than a human. The disclosure must be clear and happen at the point of interaction.

The one exception: disclosure is not required where it is obvious from the circumstances and context of use that the user is interacting with an AI. This “obviousness” bar is higher than most teams assume. A chatbot plainly badged as a virtual assistant may qualify; an AI voice designed to pass as a human call-centre agent almost certainly will not.

2. Synthetic content generation — Article 50(2), the “marking rule”

Providers of AI systems that generate or manipulate synthetic audio, image, video or text must ensure outputs are marked in a machine-readable format and detectable as artificially generated. This is a technical, provenance-focused duty — about enabling detection tools to verify machine origin, through watermarking, metadata and similar techniques.

It does not apply where the AI performs only an assistive or standard editing function that does not substantially alter the input. The practical test that has emerged: did the AI change the meaning, structure or substance? Spell-checking, grammar correction and formatting are generally outside the rule; generating or rewriting the substance is inside it. The precise technical standards are being finalised through a dedicated Code of Practice and EU standardisation work.

3. Emotion recognition & biometric categorisation — Article 50(3)

Deployers of emotion recognition or biometric categorisation systems must inform the individuals exposed to them that such a system is operating, before or at the time of exposure. The exception is use permitted by law to detect, prevent or investigate criminal offences, subject to safeguards. A separate GDPR lawful basis is still required for the underlying personal-data processing.

4. Deepfakes & AI-generated public-interest text — Article 50(4)

This is the most scrutinised obligation, and it binds deployers — often the publisher, brand, agency or comms team creating and distributing content. It covers two cases.

Deepfakes. A deepfake is defined in Article 3(60) as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful. Deployers must disclose that such content has been artificially generated or manipulated. The contexts this lands on include advertising with synthetic depictions of real or realistic-looking people, political communication, influencer and brand content, corporate communications, product demonstrations, and journalism using synthetic re-enactments.

AI-generated text on matters of public interest. Where AI-generated text is published to inform the public on matters of public interest, its artificial origin must be disclosed. This is narrower than it sounds: most internal documentation, marketing copy and product literature falls outside it. Public-health communications, scientific commentary for general audiences, journalism on public-interest topics and certain corporate statements on public issues can fall inside it.

A crucial clarification from the draft guidelines: the deepfake labelling duty does not depend on any intention to deceive. Content resembling a real-looking person must be labelled even if no deception was intended — and even if the depicted person is entirely fictitious but natural-looking. Conversely, clearly fantastical or physically impossible content — dragons, humans flying unaided — falls outside the deepfake definition.

The exemptions that genuinely apply

Article 50 contains several real exceptions. Knowing them precisely is the difference between over-labelling everything and complying intelligently.

• Obviousness (50(1)). No chatbot disclosure where it is already obvious to a reasonable person that they are interacting with AI.
• Assistive / standard editing (50(2)). No marking where AI only assists or performs standard editing without substantially altering the input.
• Artistic, creative, satirical or fictional works (50(4)). For deepfakes that are evidently part of such a work, disclosure is attenuated — it must still appear, but in a way that does not hamper the display or enjoyment of the work (for example, in the credits rather than across the frame). Third-party rights such as intellectual property and personality/privacy rights must still be safeguarded. Note the limit: an AI-generated video of a real politician saying things they never said is not rescued by “artistic” framing.
• Human editorial responsibility (50(4) text). The public-interest-text duty does not apply where the content has undergone genuine human review or editorial control and a natural or legal person holds editorial responsibility. The draft Code is explicit that this must be substantive — a documented editorial workflow with identified responsible people, not a cursory glance before hitting publish.
• Law enforcement. Specific carve-outs apply where use is authorised by law to detect, prevent, investigate or prosecute criminal offences.
• Personal and research contexts. Purely personal, non-professional activity and scientific research and development are outside scope.

Important interaction: even where a deployer benefits from an exemption (say, editorial responsibility for an article), the provider’s machine-readable marking obligation under 50(2) can still apply to the underlying generated content. The exemptions are obligation-specific, not a blanket pass.

How a disclosure must actually be made

The Commission’s guidelines are specific about quality of disclosure, and equally specific about what does not count. The notification must be:

• Provided at the latest at the first interaction or exposure — for a chatbot, before or at the very start of the conversation; for an audio or video clip, at the beginning. In sensitive contexts a one-time disclosure may not be enough and may need to be repeated.
• Clear and distinguishable, in plain language, and conforming to applicable accessibility requirements.
• Perceivable by the user at the point of interaction — for deployer labelling this means user-facing signals: visible labels or icons for images and video, audible disclaimers for audio, prominent flags for public-interest text.

What the guidelines say does not satisfy the obligation: disclosures buried in terms and conditions, URLs or documentation; machine-readable marking alone (it is not perceivable by users); vague signals like a generic “assistant” label; and technical descriptions such as “this system uses LLMs” that do not tell the user what the system actually is for them.

It applies even if you are not in the EU

Article 50’s reach is extraterritorial. Providers established outside the EU are in scope where they place AI systems on the EU market, or where the system’s output is used in the EU. Deployers outside the EU are in scope where the output is used in the EU. In plain terms: a non-EU chatbot serving EU customers, or non-EU software generating content seen by EU users, is caught. Geography is not a shield.

Penalties for getting it wrong

Breaching the Article 50 transparency obligations sits in the AI Act’s middle penalty tier: fines of up to €15 million or 3% of total worldwide annual turnover, whichever is higher. EU institutions, bodies and agencies face a separate ceiling of up to €750,000. These are maximums, and actual fines depend on the nature of the infringement — but the figures make clear transparency is not treated as a minor formality. (For context, the Act’s top tier, for prohibited practices, reaches €35 million or 7% of turnover.)

The Code of Practice — voluntary, but the de facto benchmark

To help organisations comply, the AI Office has been developing a Code of Practice on the transparency of AI-generated content, drafted by independent experts. Adherence is formally voluntary. In practice, it is widely expected to become the benchmark against which compliance with the marking and labelling obligations (50(2) and 50(4)) is assessed. Signing up to a Code the AI Office considers adequate is a straightforward way to demonstrate compliance; non-signatories can expect more information requests and will need to justify their alternative measures, for example through a gap analysis against the Code. The Code also stresses that labelling cannot rely on automation alone — it needs human review and editorial processes to function.

What to do before 2 August 2026

• Inventory every AI touchpoint. Web chat, voice assistant, phone IVR, kiosk, API-served responses, image and copy generators, any synthetic media. You cannot label what you have not mapped.
• Classify your role per touchpoint. Provider, deployer, or both — and which of the four obligations attaches.
• Implement perceivable disclosures at first interaction. Chatbot identification, AI-content labels, deepfake disclosures — visible/audible and contemporaneous, never hidden in legal text.
• Build machine-readable marking pipelines. For synthetic content you provide, assess whether your systems support machine-readable marking and detectable provenance; track the Code and standards as they finalise.
• Document editorial workflows. If you will rely on the editorial-responsibility exemption for public-interest text, set up a real, documented workflow with named responsible people.
• Review AI supply-chain contracts. Responsibility does not rest solely with the original model provider; clarify who marks, who labels, and who is liable across the chain.
• Keep records. Document your assessment and the measures taken — it is your first line of defence in any information request.

Frequently asked questions

Does Article 50 apply to my chatbot if it is obviously a bot?

If it is genuinely obvious from context that a user is interacting with AI, the 50(1) disclosure may not be required. But “obvious” is judged by a reasonable user, not by your assumptions — when in doubt, disclose. It is cheap insurance.

Do I need to label AI-generated marketing copy?

The public-interest-text obligation (50(4)) does not generally capture ordinary marketing copy, which is not “informing the public on matters of public interest.” However, if you are the provider of the generation system, the machine-readable marking duty (50(2)) can still apply to the output, and labelling AI involvement is increasingly treated as best practice regardless.

What if AI only lightly edits human-written text?

Minor assistive edits — spell-check, grammar, formatting — are generally exempt. The test is whether the AI changed the meaning, structure or substance. If it only corrected typos, it is not “generation or manipulation” for these purposes.

Are artistic or satirical deepfakes exempt?

Not fully. For works that are evidently artistic, creative, satirical or fictional, the disclosure is attenuated — it must still appear but in a way that does not spoil the work (for example, in the credits). It does not excuse realistic depictions of real people in misleading ways.

Does Article 50 apply to companies outside the EU?

Yes, where the AI system is placed on the EU market or its output is used in the EU. A US or UK provider or deployer serving EU users is in scope.

When exactly does Article 50 take effect?

2 August 2026 under the law in force as of June 2026. The Digital Omnibus reform package largely leaves this date intact, but as it was not yet formally adopted at the time of writing, confirm the current position before relying on any date.

The bottom line

Article 50 is the AI Act obligation most likely to affect the most organisations, precisely because it does not hinge on a high-risk classification. If your product includes a chatbot, generates images or text, uses emotion recognition, or creates deepfakes, you have disclosure and labelling duties from 2 August 2026. The Code of Practice is voluntary; Article 50 is not. The organisations that arrive prepared will be the ones that started the inventory now — while the standards are still settling and there is time to build transparency in by design.