EU AI Act Article 50: Transparency and Labelling

Most EU AI Act coverage is about high-risk systems, whose obligations the Digital Omnibus pushed back to December 2027. Article 50 is different: it has applied since 2 August 2026, and it reaches far more companies, because it doesn’t depend on your system being high-risk. If you run a customer-facing chatbot, generate images, audio, video or text with AI, or publish AI-written content, Article 50 is very likely already your nearest live obligation.

There is also one date still ahead. Generative AI systems that were already on the market before 2 August 2026 have until 2 December 2026 to meet the machine-readable marking requirement in Article 50(2). This guide covers what each paragraph requires, who carries which duty, what the Commission’s Code of Practice and guidelines add, and what to do before December.

Want the short answer first? The free Article 50 check tells you which duties apply to your system, and whether the December deadline does, in about a minute.

The Four Duties at a Glance

Article 50 sets out four separate transparency duties. Two fall on providers (the company that builds the system or places it on the market) and two on deployers (the company that uses it). If you’re unsure which role you hold, see our provider vs deployer guide — many companies are both.

Article 50(5) adds a rule for all four: the information must be given “in a clear and distinguishable manner at the latest at the time of the first interaction or exposure”, and it must meet applicable accessibility requirements.

Article 50(1): Telling People They’re Talking to AI

Providers must design systems that interact directly with people — chatbots, voice assistants, support agents — so that users are informed they are interacting with an AI system. The exception is where that is obvious to “a natural person who is reasonably well-informed, observant and circumspect”, given the context.

The Commission’s Article 50 guidelines, published on 20 July 2026, set a practical bar:

• Disclosure has to be visible at first interaction. A line in the terms of service, or a generic label like “assistant”, is not enough.
• AI agents are covered. An agent that interacts with people should disclose both that it is an AI and on whose behalf it is acting.
• Don’t lean on the “obvious” exception. It is read narrowly, and the bar is higher when your users may include children, older people or people with disabilities.

Article 50(2): Marking AI-Generated Content

Providers of systems that generate synthetic audio, image, video or text — including general-purpose AI systems — must ensure outputs are “marked in a machine-readable format and detectable as artificially generated or manipulated.” The technical solution must be effective, interoperable, robust and reliable “as far as this is technically feasible.”

There are two carve-outs in the text: systems that perform an assistive function for standard editing, and systems that don’t substantially alter the input data or its meaning.

This is the paragraph with the 2 December 2026 deadline for systems that were already on the market before 2 August 2026, set by the Omnibus amendment to Article 111. New systems placed on the market since 2 August 2026 have had to comply from day one.

What the Code of Practice expects

The Commission published the final Code of Practice on marking and labelling AI-generated content on 10 June 2026, and the Commission and the AI Board have since assessed it as adequate. Signing it is voluntary, but it’s the most straightforward way to show how you meet Article 50(2). According to law-firm analyses of the final text, it expects providers to:

• Use more than one marking technique. No single method is considered sufficient today. The typical pairing is signed, tamper-evident metadata plus an imperceptible watermark.
• Offer a way to detect your marks, in principle free of charge, with free access for authorities, researchers and fact-checkers.
• Make detection interoperable with other providers’ marks by 2 February 2027.

If you don’t sign the Code, you can still comply, but you’ll need to show your own measures achieve the same result. In practice that means a documented gap analysis against the Code.

Article 50(3): Emotion Recognition and Biometric Categorisation

Deployers of emotion recognition or biometric categorisation systems must inform the people exposed to them, and process their personal data in line with the GDPR.

A notice doesn’t make the system lawful. Emotion recognition in the workplace and in education is prohibited under Article 5, except for medical or safety reasons, and many biometric uses are high-risk. Check those first. Article 50(3) only covers the disclosure for uses that are otherwise permitted.

Article 50(4): Deepfakes and Public-Interest Text

Deployers carry two labelling duties:

Deepfakes. If you use AI to generate or manipulate image, audio or video content that constitutes a deep fake, you must disclose that it is artificially generated or manipulated. For evidently artistic, creative, satirical or fictional work, the duty is lighter: you disclose the existence of the generated content in a way that doesn’t spoil the work. The guidelines treat the deepfake test as objective — intent to deceive isn’t required — and most advertising doesn’t qualify for the lighter regime.

Public-interest text. If you publish AI-generated or manipulated text “with the purpose of informing the public on matters of public interest”, you must disclose it. The exception is text that has undergone human review or editorial control, where a person or organisation holds editorial responsibility. The guidelines and Code read this narrowly: a quick sign-off or spell-check doesn’t count, and you should be able to show a documented review workflow.

Two practical points from the guidance:

• The provider’s watermark doesn’t cover you. Machine-readable marking under 50(2) does not discharge a deployer’s duty to show a visible label under 50(4).
• Use the EU icon. The Code includes standard EU icons for AI-generated and AI-modified content. For audio-only content, a spoken or written disclaimer works instead.

Content generated or published before 2 August 2026 doesn’t need to be labelled retroactively.

What Article 50 Doesn’t Change

Article 50 sits alongside, not instead of, everything else. Article 50(6) says it doesn’t affect the high-risk requirements in Chapter III or transparency rules in other EU or national law. A chatbot used to screen job applicants needs its Article 50(1) disclosure and the full high-risk treatment from December 2027.

Breaching Article 50 falls in the Act’s middle fine tier: up to €15 million or 3% of worldwide annual turnover, whichever is higher, with lower caps for SMEs. See EU AI Act fines for how the tiers work.

What to Do Before 2 December 2026

1. List every AI system that faces people or produces content. That includes chatbots, support agents, voice tools, image and video generators, and AI writing tools — including features embedded in vendor software. Your AI system inventory is the place to record this.
2. Assign the role for each one. Provider duties (50(1) and 50(2)) sit with whoever builds or markets the system. Deployer duties (50(3) and 50(4)) sit with whoever uses it. If you put your own name on a vendor’s model, you may be the provider.
3. Check the date each generative system was placed on the market. Before 2 August 2026, the marking deadline is 2 December 2026. On or after, it already applies.
4. Ask your generative AI vendors how they mark output and whether they’ve signed the Code. If you’re the provider, decide whether to sign it or document your own gap analysis.
5. Put a visible label on deepfakes and public-interest text, using the EU icon where possible. If you rely on the editorial exception, write the review workflow down.
6. Fix chatbot disclosure so it appears at the first interaction, in plain language, and covers agents acting on someone’s behalf.

How Aikraft Helps

Aikraft’s classification questionnaire asks the questions that trigger Article 50: does the system interact with people, generate synthetic content, or infer emotions? The answers are recorded in the same inventory as your high-risk and prohibited-practice checks, so transparency duties aren’t tracked in a separate spreadsheet. To see which transparency duties apply to a system right now, run the free Article 50 check; for the rest of the Act, take the risk quiz. Neither needs a signup.

This guide reflects Article 50 as published in Regulation (EU) 2024/1689, the Digital Omnibus (Regulation (EU) 2026/1744), the Commission’s Code of Practice (10 June 2026) and its Article 50 guidelines (20 July 2026). The Code and guidelines are not legally binding. This is a summary, not legal advice.