AI Act: what Article 50 changes on your screens

When we audit AI screens, the same scene keeps coming up. The legal team has read the regulation, the product team has shipped the feature, and nobody has looked at the precise spot where the two meet: the chatbot bubble, the label under an image, the first message of a call. Yet that is exactly where Article 50 places its obligations.

This article sets out what the text actually requires, how to translate it into screens, and the mistakes we see most often. It doesn’t replace a lawyer’s advice, but it should help you go into that meeting with the right questions.

Where the timeline stands

The European regulation on artificial intelligence (Regulation 2024/1689) entered into force on 1^st August 2024, with phased application:

• 2 February 2025 : ban on certain practices (social scoring, manipulation, etc.) and an AI literacy obligation for teams using these systems.
• 2 August 2025 : obligations for providers of general-purpose AI models, the ones who build the large models.
• 2 August 2026 : general application of the regulation, including Article 50 on transparency.
• 2 August 2027 : obligations for high-risk AI embedded in already regulated products (medical devices, machinery, toys…).

In late 2025, the Commission proposed a simplification package, the “digital omnibus”, which notably plans to postpone some high-risk AI obligations and grant extra time for certain technical marking requirements. These adjustments are still under discussion: check the status of the text when you read this. But no serious scenario does away with the principle behind Article 50: people must know when they are dealing with an AI.

What Article 50 actually says

Article 50 contains four obligations. They don’t all fall on the same parties, and that’s the first thing to clarify within your team. The regulation distinguishes the provider (who develops the system and places it on the market under their own name) from the deployer (who uses it in the course of their business). A company that plugs a large model’s API into its application and offers it under its own brand is often both, depending on the feature.

1. Disclose that people are talking to an AI. A system designed to interact directly with people must inform them that they are interacting with an AI, unless this is obvious to a reasonably attentive person given the context. This is the obligation that covers chatbots, voice assistants and agents that reply by email.
2. Mark generated content. Systems that generate text, images, audio or video must mark their output in a machine-readable format, so it can be detected as artificial. This obligation mainly targets the providers of generation systems. It doesn’t apply to writing-assistance features that don’t substantially alter the user’s text.
3. Give notice before analysing emotions or biometric characteristics. If you deploy an emotion recognition or biometric categorisation system, the people exposed to it must be informed.
4. Disclose deepfakes and certain texts. Images, audio or video that deceptively imitate real people, places or events must be disclosed as generated. The same goes for generated texts published to inform the public on matters of public interest, unless they have been reviewed by a person who takes editorial responsibility for them.

The fifth paragraph of the article is the one people forget, yet it’s the most concrete for a product team: the information must be provided in a clear and distinguishable manner at the latest at the time of the first interaction or exposure, and it must meet accessibility requirements.

What it means, screen by screen

The support chatbot

This is the most common case, and the most poorly handled. The principle is simple: before or at the first message, the person must know they are not talking to a human agent. In practice, that means a sentence in the bot’s first message (“I’m the automated assistant for…”), a name that doesn’t pass itself off as a human first name, and a persistent label in the conversation header.

What isn’t enough: a mention in the terms and conditions, a sparkle icon with no text, an avatar of a smiling person called “Julie”. If your bot is called Julie, the question is no longer legal, it’s one of good faith.

Voice assistants and phone systems

A realistic synthetic voice makes the “it’s obvious” exception practically unusable: the notice must be spoken at the start of the call, before any question. Also plan for the moment the call is handed over to a human, and say so. People adjust how they speak depending on whether they’re addressing a machine or a person, and that is precisely what the text aims to protect.

Generated content in your product

If your application generates images, product visuals or voices, technical marking (invisible watermarks, provenance metadata) generally falls to the model provider. But if you are that provider, or if you transform a model’s output before delivering it, the question comes back to you. Ask your provider what they mark, how, and whether your processing pipeline (compression, cropping, conversion) preserves that marking. It’s as much an architecture question as a design one.

On the interface side, nothing stops you from going beyond the obligation: a small “Generated image” label on AI-created visuals avoids a lot of misunderstandings with your customers, especially when those visuals are shared outside your product.

AI-written text

This is where a quick reading of the text is misleading. The obligation to disclose generated text applies to texts published to inform the public on matters of public interest, and it no longer applies if a person has reviewed them and takes responsibility for them. A product page, a sales email or a support reply drafted with a model’s help doesn’t automatically fall into this category. On the other hand, a media outlet or a public authority publishing generated articles without review is directly concerned.

The “it’s obvious” exception: handle with care

The text exempts disclosure when the artificial nature is obvious “from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use”. We often see teams rely on this sentence to display nothing at all.

Our advice is the opposite: assume nothing is obvious. What’s obvious to your team, who designed the product, isn’t obvious to an elderly person contacting their insurer on a Sunday evening, or to someone using a screen reader who can’t see the purple icon that’s supposed to say it all. A sentence costs three seconds of reading. A dispute over whether an interface was “obvious” costs a great deal more.

A notice everyone must be able to perceive

Article 50 requires the information to meet accessibility requirements. Teams often discover this late, and it has very practical consequences:

• An icon alone isn’t enough: it needs a text equivalent read by screen readers.
• The notice must be perceivable without relying on colour, with sufficient contrast and a legible size.
• In a chatbot, the first message must be announced by assistive technologies, which requires a properly marked-up conversation area.
• For voice, the spoken notice must be backed by a written equivalent when a screen is available.

The good news: if you’ve already worked on compliance with the European Accessibility Act, in force since June 2025 for many digital services, you’re halfway there. The two workstreams are best tackled together.

The mistakes we see most often

1. The hidden notice. In a footer, a tooltip or the terms of service. The text requires clear information at the time of the first interaction, not information available somewhere.
2. The fake human. A first name, a stock photo, “let me look into that for you” with a typing indicator tuned to mimic a human. Anything that sustains the illusion goes against the spirit of the text.
3. The sparkle as the only signal. The ✨ icon has become the symbol for AI in many products, but it has no universally shared meaning, and it is silent to a screen reader.
4. The silent handover. A human agent takes over the conversation, or the other way round, without anyone saying so.
5. Marking lost along the way. The model marks its images, then your pipeline compresses them and strips the metadata.

The checklist to run before going to production

• Every AI feature is listed, with your role for each: provider, deployer, or both.
• Every direct interaction with an AI is disclosed from the first exchange, in plain words.
• The assistant’s name and avatar don’t suggest a real person.
• Handovers between AI and human are announced in both directions.
• Notices are readable by screen readers and don’t rely on colour or an icon alone.
• You know what your model provider marks, and your processing pipeline preserves it.
• Generated content imitating real people or places is disclosed.
• Informational texts published without human review are identified as generated.

What Article 50 doesn’t require, but your users expect

Article 50 is a transparency obligation, not a quality one. For a system that isn’t high-risk, it doesn’t require you to provide human oversight, explain every suggestion or display a confidence level. Yet these are precisely the elements that determine whether an AI feature gets adopted or ignored.

Saying “this is an AI” is the minimum. Saying “here’s why I’m suggesting this, here’s my source, and you decide” is what builds trust. We go into these interface choices in another article: why your users don’t trust your AI feature.

This article presents our reading of the regulation for product teams. It does not constitute legal advice: for a specific case, have your analysis validated by a specialised lawyer.

Frequently asked questions

Does Article 50 of the AI Act apply to small businesses?

Yes. The transparency obligations apply regardless of company size. Only the penalty cap is adjusted: for SMEs and start-ups, the lower of €15 million and 3% of worldwide turnover applies.

Do you have to disclose text written with the help of AI?

Not systematically. The obligation covers texts published to inform the public on matters of public interest, and it no longer applies if a person has reviewed them and takes editorial responsibility. A reviewed product page or email is not covered on that basis.

Is a sparkle icon enough to signal AI?

No. The information must be clear, distinguishable and accessible. An icon without text has no universally shared meaning and isn’t perceived by people using a screen reader.

When must the user be informed?

At the latest at the first interaction or first exposure to the content. In a chatbot, that means from the very first message, not after several exchanges.